Introduction to Formal Concept Analysis and Its

Applications in Information Retrieval and

Related Fields

Dmitry I. Ignatov

National Research University Higher School of Economics, Moscow

[email protected]

Abstract. This paper is a tutorial on Formal Concept Analysis (FCA)

arXiv:1703.02819v1 [cs.IR] 8 Mar 2017

and its applications. FCA is an applied branch of Lattice Theory, a math-

ematical discipline which enables formalisation of concepts as basic units
of human thinking and analysing data in the object-attribute form. Orig-
inated in early 80s, during the last three decades, it became a popular
human-centred tool for knowledge representation and data analysis with
numerous applications. Since the tutorial was specially prepared for RuS-
SIR 2014, the covered FCA topics include Information Retrieval with
a focus on visualisation aspects, Machine Learning, Data Mining and
Knowledge Discovery, Text Mining and several others.

Keywords: Formal Concept Analysis, Concept Lattices, Information

Retrieval, Machine Learning, Data Mining, Knowledge Discovery, Text
Mining, Biclustering, Multimodal Clustering

1 Introduction

According to [1], “information retrieval (IR) is finding material (usually doc-

uments) of an unstructured nature (usually text) that satisfies an information
need from within large collections (usually stored on computers).” In the past,
only specialized professions such as librarians had to retrieve information on a
regular basis. These days, massive amounts of information are available on the
Internet and hundreds of millions of people make use of information retrieval
systems such as web or email search engines on a daily basis. Formal Concept
Analysis (FCA) was introduced in the early 1980s by Rudolf Wille as a mathe-
matical theory [2,3] and became a popular technique within the IR field. FCA is
concerned with the formalisation of concepts and conceptual thinking and has
been applied in many disciplines such as software engineering, machine learning,
knowledge discovery and ontology construction during the last 20-25 years. In-
formally, FCA studies how objects can be hierarchically grouped together with
their common attributes.
The core contributions of this tutorial from IR perspective are based on our
surveys [4,5,6] and experiences in both fields, FCA and IR. In our surveys we
visually represented the literature on FCA and IR using concept lattices as well
2 Dmitry I. Ignatov

as several related fields, in which the objects are the scientific papers and the
attributes are the relevant terms available in the title, keywords and abstract of
the papers. You can see an example of such a visualisation in Figure 1 for papers
published between 2003 and 2009. We developed a toolset with a central FCA
component that we used to index the papers with a thesaurus containing terms
related to FCA research and to generate the lattices. The tutorial also contains
a partial overview of the papers on using FCA in Information Retrieval with a
focus on visualisation.

Fig. 1. The lattice diagram representing collection of 702 papers on FCA including 103
papers on FCA and IR (2003-2009).

In 2013 European Conference on Information Retrieval [7] hosted a thematic

workshop FCA meets IR which was devoted to two main issues:

– How can FCA support IR activities including but not limited to query anal-
ysis, document representation, text classification and clustering, social net-
work mining, access to semantic web data, and ontology engineering?
– How can FCA be extended to address a wider range of IR activities, possibly
including new retrieval tasks?
 Introduction to Formal Concept Analysis and Its Applications in IR 3

Claudio Carpineto delivered an invited lecture at the workshop – “FCA and

IR: The Story So Far”. The relevant papers and results presented there are also
discussed in the tutorial.
Since the tutorial preparations were guided by the idea to present the content
at a solid and comprehensible level accessible even for newcomers, it is a balanced
combination of theoretical foundations, practice and relevant applications. Thus
we provide intro to FCA, practice with main tools for FCA, discuss FCA in
Machine Learning and Data Mining, FCA in Information Retrieval and Text
Mining, FCA in Ontology Modeling and other selected applications. Many of
the used examples are real-life studies conducted by the course author.
The target audience is Computer Science, Mathematics and Linguistics stu-
dents, young scientists, university teachers and researchers who want to use FCA
models and tools in their IR and data analysis tasks.
The course features five parts. Each part is placed in a separate section and
contains a short highlight list to ease the navigation within the material. An
archive with supplementary files for exercises and examples is available at 1 .
Section 2 contains introduction to FCA and related notions of Lattice and Or-
der Theory. In Section 3, we describe selected FCA tools and provide exercises.
Section 4 provides an overview of FCA-based methods and applications in Data
Mining and Machine Learning, and describes an FCA-based tool for supervised
learning, QuDA (Qualitative Data Analysis). Section 5 presents the most rele-
vant part of the course, FCA in Information Retrieval and Text Mining. Penul-
timate Section 6 discusses FCA in Ontology Modeling and gives an example of
FCA-based Attribute Exploration technique on building the taxonomy of trans-
portation means. Section 7 concludes the paper and briefly outlines prospects
and limitations of FCA-based models and techniques.

2 Introduction to FCA
Even though that many disciplines can be dated back to Aristotles time, more
closer prolegomena of FCA can be found, for example, in the Logic of Port Royal
(1662)[8], an old philosophical concept logic, where a concept was treated as a
pair of its extent and its intent (yet without formal mathematical apparatus).
Being a part of lattice theory, concept lattices are deeply rooted in works
of Dedekind, Birkgoff [9] (Galois connections and “polarities”), and Ore [10]
(Galois connections), and, later, on Barbut&Monjardet [11] (treillis de Galois,
i.e. Galois lattices).
In fact, the underlying structure, Galois connection, has a strong impact in
Data Analysis[12,13,14,15].
In this section, we mainly reproduce basic definitions from Ganter&Wille’s
book on Formal Concept Analysis [3]. However, one can find a good introductory
material, more focused on partial orders and lattices, in the book of Davey and
Priestly [16]. An IR-oriented reader may also find the following books interesting
and helpful [15,17].
1
http://bit.ly/RuSSIR2014FCAtut
4 Dmitry I. Ignatov

There were several good tutorials with notes in the past, for example, a basic
one [18] and more theoretical with algorithmic aspects [19].
We also refer the readers to some online materials that might be suitable for
self-study purposes 2 ,3 ,4 .
A short section summary:

– Binary Relations, Partial Orders, Lattices, Line (Hasse) Diagram.

– Galois Connection, Formal Context, Formal Concept, Concept Lattice.
– Concept Lattice drawing. Algorithms for concept lattices generation (naı̈ve,
Ganter’s algorithm, Close-by-One).
– Attribute Dependencies: implications, functional dependencies. Armstrong
Rules. Implication bases (Stem Base, Generator base).
– Many-valued contexts. Concept scaling.

2.1 Binary Relations, Partial Orders, Lattices, Hasse Diagram

The notion of a set is fundamental in mathematics. In what follows, we consider

only finite sets of objects.

Definition 1. A binary relation R between two sets A and B is a set of all

pairs (a, b) with a ∈ A and b ∈ B., i.e., a subset of their Cartesian product
A × B, the set of all such pairs.

Sometimes it is convenient to write aRb instead (a, b) ∈ R for brevity. If

A = B then R ⊆ A × A is called a binary relation on the set A.

Definition 2. A binary relation R on a set A is called a partial order relation

(or shortly a partial order), if it satisfies the following conditions for all elements
a, b, c ∈ A:

1. aRa (reflexivity)
2. aRb and a 6= b =⇒ not aRb (antisymmetry)
3. aRb and bRc =⇒ aRc (transitivity)

We use symbol ≤ for partial order, and in case a ≤ b and a 6= b we write

a ≤ b. We read a ≤ b as “a is less of equal to b”. A partially ordered set (or
poset) is a pair (P, ≤), where P is a set and ≤ is an partial order on P .

Definition 3. Given a poset (P, ≤), an element a is called a lower neighbour

of b, if a ≤ b and there is no such c fulfilling a ≤ c ≤ b. In this case, b is also
an upper neighbour of a, and we write a ≺ b.
2
http://www.kbs.uni-hannover.de/~jaeschke/teaching/2012w/fca/
3
http://www.upriss.org.uk/fca/fcaintro.html
4
http://ddll.inf.tu-dresden.de/web/Introduction_to_Formal_Concept_
Analysis_(WS2014)/en
 Introduction to Formal Concept Analysis and Its Applications in IR 5

Every finite ordered poset (P, ≤) can be depicted as a line diagram (many
authors call it Hasse diagram). Elements of P are represented by small circles
in the plane. If a ≤ b, the circle corresponding to a is depicted higher than the
circle corresponding to b, and the two circles are connected by a line segment.
One can check whether some a ≤ b if there is an ascending path from b to a in
the diagram.

Example 1. The poset P is given by its incidence cross-table, where × in a cell

means that the corresponding pair of row and column elements x and y are
related as follows x ≤ y.

≤ a b c d e
a ×
b ×××

c ×
d ×××××
e ×

The graph of P .
The line diagram of P .

Definition 4. Let (P, ≤) be a poset and A a subset of P . A lower bound of A

is an element l of P with l ≤ A for all a ∈ A. An upper bound of A is defined
dually. If there is a largest element in the set ofVall lower bounds of A, it is called
the infimum of A and is denoted by inf A or A. Dually, if there is a smallest
element inW the set of all upper bounds, it is called supremum and denoted by
supA or A.

For A = {a, b} we write x∧y for inf A and x∨y for supA. Infimum and supremum
are also called meet and join.

Definition 5. A poset L = (L, ≤) is a lattice, if for any two elements a and b in

L the supremum a ∨ b andWthe infimum a ∧ b always V exist. L is called a complete
lattice, if the supremum X and the infimum X exist forWany subset A of L.
For every complete lattice L there exist its largest element, L, called the unit
element of the lattice, denoted by 1L . Dually, the smallest element 0L is called
the zero element.

Example 2. In Fig. 2 there are the line diagrams of the poset P , which is not
a lattice, and the lattice L. It is interesting that P has its largest and smallest
elements, 1P and 0P ; the pair of its elements, s and t, has its infumum, s∨t = 0P ,
but there is no a supremum for it. In fact, p, t does not have a smallest element
in the set of all its upper bounds.

6 Dmitry I. Ignatov
1P 1L

q p

v∨w

t s v w

0P s ∧ t 0L v ∧ w

Fig. 2. The line diagrams of the order, which is not a lattice (left), and the order,
which is a lattice (right)

3 Galois Connection, Formal Context, Formal Concept,

Concept Lattice
Definition 6. Let ϕ : P → Q and ψ : Q → P be maps between two posets
(P, ≤) and (Q, ≤). Such a pair of maps is called a Galois connection between
the ordered sets if:
1. p1 ≤ p2 ⇒ ϕp1 ≥ ϕp2
2. q1 ≤ q2 ⇒ ψq1 ≥ ψq2
3. p ≤ ψϕp ⇒ q ≤ ϕψq.
Exercise 1. Prove that a pair (ϕ, ψ) of maps is a Galois connection if and only
if p ≤ ψq ⇔ q ≤ ψp. 
Exercise 2. Prove that for every Galois connection (ϕ, ψ)
ψ = ψϕψ and ϕ = ϕψϕ.

Definition 7. A formal context K = (G, M, I) consists of two sets G and M
and a relation I between G and M . The elements of G are called the objects
and the elements of M are called the attributes of the context. The notation
gIm or (g, m) ∈ I means that the object g has attribute m.
Definition 8. For A ⊆ G, let

A0 := {m ∈ M |(g, m) ∈ I for all g ∈ A}

and, for B ⊆ M , let

B 0 := {g ∈ G|(g, m) ∈ I for all m ∈ B}.

These operators are called derivation operators or concept-forming
operators for K = (G, M, I).
 Introduction to Formal Concept Analysis and Its Applications in IR 7

Proposition 1. Let (G, M, I) be a formal context, for subsets A, A1 , A2 ⊆ G

and B ⊆ M we have
1. A1 ⊆ A2 iff A02 ⊆ A01 ,
2. A ⊆ A00 ,
3. A = A000 (hence, A0000 = A00 ),
4. (A1 ∪ A2 )0 = A01 ∩ A02 ,
5. A ⊆ B 0 ⇔ B ⊆ A0 ⇔ A × B ⊆ I.
Similar properties hold for subsets of attributes.
Exercise 3. Prove properties of operator (·)0 from proposition 1. 
Definition 9. A closure operator on set G is a mapping ϕ : 2G → 2G with
the following properties:
1. ϕϕX = ϕX (idempotency)
2. X ⊆ ϕX (extensity)
3. X ⊆ Y ⇒ ϕX ⊆ ϕY (monotonicity)
For a closure operator ϕ the set ϕX is called closure of X.
A subset X ⊆ G is called closed if ϕX = X.
Exercise 4. Let (G, M, I) be a context, prove that operators
(·)00 : 2G → 2G , (·)00 : 2M → 2M
are closure operators. 
Definition 10. A formal concept of a formal context K = (G, M, I) is a
pair (A, B) with A ⊆ G, B ⊆ M , A0 = B and B 0 = A. The sets A and B
are called the extent and the intent of the formal concept (A, B), respectively.
The subconcept-superconcept relation is given by (A1 , B1 ) ≤ (A2 , B2 ) iff
A1 ⊆ A2 (B1 ⊆ B2 ).
This definition says that every formal concept has two parts, namely, its ex-
tent and intent. This follows an old tradition in the Logic of Port Royal (1662),
and is in line with the International Standard ISO 704 that formulates the fol-
lowing definition: “A concept is considered to be a unit of thought constituted
of two parts: its extent and its intent.”
Definition 11. The set of all formal concepts of a context K together with the
order relation I forms a complete lattice, called the concept lattice of K and
denoted by B(K).
Example 3. The context with four geometric figures and four attributes is below.
G\M a b c d
Objects: Attributes:
1 × ×
1 – equilateral triangle, a – has 3 vertices,
2 × × 2 – rectangle triangle, b – has 4 vertices, 
3 ×× 3 – rectangle, c – has a direct angle,
4 – square. d – equilateral.
4 ×××
8 Dmitry I. Ignatov

({1,2,3,4},∅)

({1,4},{d}) ({2,3,4},{c}) ({1,2},{a})

({3,4},{b,c})

({4},{b,c,d}) ({1},{a,d}) ({2},{a,c})

(∅,M)

Fig. 3. The line diagram of the concept lattice for the context of geometric figures

Definition 12. For every two formal concepts (A1 , B1 ) and (A2 , B2 ) of a cer-
tain formal context their greatest common subconcept is defined as follows:

(A1 , B1 ) ∧ (A2 , B2 ) = (A1 ∩ A2 , (B1 ∪ B2 )00 ).

The least common superconcept of (A1 , B1 ) and (A2 , B2 ) is given as

(A1 , B1 ) ∨ (A2 , B2 ) = ((A1 ∪ A2 )00 , B1 ∩ B2 ).

We say supremum instead of “least common superconcept”, and instead of

“greatest common subconcept” we use the term infimum.

(A1 , B1 ) ∨ (A2 , B2 )

(A2 , B2 ) (A1 , B1 )

(A1 , B1 ) ∧ (A2 , B2 )

Fig. 4. Supremum and infimum of two concepts

It is possible to define supremum and infumum operations for an arbitrary

set of concepts of a certain context. This is done in the first part of Theorem 1.

Definition 13. A subset X ⊆ L of lattice (L, ≤) is called supremum-dense

if any lattice element v ∈ L can be represented as
_
v = {x ∈ X | x ≤ v}.

Dually for infimum-dense subsets.

 Introduction to Formal Concept Analysis and Its Applications in IR 9

The Basic Theorem of Formal Concept Analysis below defines not only supre-
mum and infimum of arbitrary sets of concepts; it also answer the question
whether concept lattices have any special properties. In fact, the answer is “no”
since every concept lattice is (isomorphic to some) complete lattice. That is one
can compose a formal context with objects G, attributes M and binary relation
I ⊂ G × M such that the original complete lattice is isomorphic B(G, M, I).
Even though the theorem does not reply how such a context can be built, but
rather describes all possibilities to do this.

Theorem 1. Basic Theorem of Formal Concept Analysis ([Wille 1982],[Gan-

ter, Wille 1996])
Concept lattice B(G, M, I) is a complete lattice. For arbitrary sets of formal
concepts
{(Aj , Bj ) | j ∈ J} ⊆ B(G, M, I)
their infimum and supremum are given in the following way:
^ \ [
(Aj , Bj ) = ( Aj , ( Bj )00 ),
j∈J j∈J j∈J

_ [ \
(Aj , Bj ) = (( Aj )00 , Bj ).
j∈J j∈J j∈J

A complete lattice L is isomorphic to a lattice B(G, M, I) iff there are mappings

γ : G → V and µ : M → V such that γ(G) is supremum-dense in L, µ(M ) is
infimum-dense in L, and gIm ⇔ γg ≤ µm for all g ∈ G and all m ∈ M . In
particular, L is isomorphic to B(L, L, ≤).

An interested reader may refer to Ganter&Wille’s book on FCA [3] for further
detailed and examples.

3.1 Concept Lattice drawing and algorithms for concept lattices

generation

One can obtain the whole set of concepts of a particular context K simply by
definition, i.e. it is enough to enumerate all subsets of objects A ⊆ G (or at-
tributes B ⊆ M ) and apply derivation operators to them. For example, for
the context from example 3 and empty set of objects, A = ∅, one may obtain
A0 = ∅0 = {a, b, c, d} = B, and then by applying (·)0 second time B 0 = ∅. Thus,
the resulting concept is (A, B) = (∅, M ).

Proposition 2. Every formal concept of a context (G, M, I) has the form (X 00 , X 0 )

for some subset X ⊆ G and the form (Y 0 , Y 00 ) for some subset Y ⊆ M . Vice
versa all such pairs of sets are formal concepts.

One may follow the naı̈ve algorithm below:

1. B(G, M, I) := ∅
10 Dmitry I. Ignatov

2. For every subset X of G, add (X 00 , X 0 ) to B(G, M, I).

Exercise 5. 1. Prove proposition 2. 2. For the context of geometric figures from

example 3 find all formal concepts. 

Since the total number of formal concept is equal to 2min(|G|,|M |) in the worst
case, this naı̈ve approach is quite inefficient even for small contexts. However,
let us assume that now we know how find concepts and we are going to build
the line diagram of a concept lattice.

1. Draw a rather small circle for each formal concept such that a circle for a
concept is always depicted higher than the all circles for its subconcepts.
2. Connect each circle with the circles of its lower neighbors.

To label concepts by attribute and object names in a concise form, we need

the notions of object and attributes concepts.

Definition 14. Let (G, M, I) be a formal context, then for each object g ∈ G
there is the object concept ({g}00 , {g}0 ) and for each attribute m ∈ M the
attribute concept is given by ({m}0 , {m}00 ).

So, if one has finished a line diagram drawing for some concept lattice, it
is possible to label the diagram with attribute names: one needs to attach the
attribute m to the circle representing the concept ({m}0 , {m}00 ). Similarly for
labeling by object names: one needs to attach each object g to the circle repre-
senting the concept ({g}00 , {g}0 ). An example of such reduced labeling is given
in Fig. 5.

d c a

3 b

4 1 2

Fig. 5. An example of reduced labeling for the lattice of geometric figures

The naı̈ve concept generation algorithm is not efficient since it enumerates

all subsets of G (or M ). For homogeneity, in what follows we reproduce the pseu-
docodes of the algorithms from [20]. There are different algorithms that compute
closures for only some subsets of G and use an efficient test to check whether the
 Introduction to Formal Concept Analysis and Its Applications in IR 11

current concept is generated first time (canonicity test). Thus, Ganter’s Next
Closure algorithm does not refer the list of generated concepts and uses little
storage space.
Since the extent of a concept defines its intent in a unique way, to obtain
the set of all formal concepts, it is enough to find closures either of subsets of
objects or subsets of attributes.
We assume that there is a linear order (<) on G. The algorithm starts by
examining the set consisting of the object maximal with respect to < (max(G)),
and finishes when the canonically generated closure is equal to G. Let A be a
currently examined subset of G. The generation of A00 is considered canonical
if A00 \ A does not contain g < max(A). If the generation of A00 is canonical
(and A00 is not equal to G), the next set to be examined is obtained from A00 as
follows:

A00 ∪ {g} \ {h|h ∈ A00 ∧ g < h}, where g = max({h|h ∈ G \ A00 }).

Otherwise, the set examined at the next step is obtained from A in a similar
way, but the added object must be less (w.r.t. <) than the maximal object in A:

A00 ∪ {g} \ {h|h ∈ A ∧ g < h}, where g = max({h|h ∈ G \ A ∧ h < max(A)}).

The pseudocode code is given in Algorithm 1 and the generation protocol of

NextClosure for the context of geometric figures is given in Table 1.

Algorithm 1 NextClosure
Input: K = (G, M, I) is a context
Output: L is the concept set
1: L := ∅, A := ∅, g := max(G)
2: while A 6= G do
3: A := A00 ∪ {g} \ {h|h ∈ A ∧ g < h}
4: if {h|h ∈ A ∧ g ≤ h} = ∅ then
5: L := L ∪ {(A00 , A0 )}
6: g := g = max({h|h ∈ G \ A00 })
7: A := A00
8: else
9: g = max({h|h ∈ G \ A ∧ h < g})
10: end if
11: end while
12: return L

The NextClosure algorithm produces the set of all concepts in time O(|G|2 |M ||L|)
and has polynomial delay O(|G|2 |M |).
We provide a simple recursive version of CbO. The algorithm generates con-
cepts according to the lectic (lexicographic) order on the subsets of G (concepts
12 Dmitry I. Ignatov

Table 1. Generation protocol of NextClosure for the context of geometric figures

g A A00 formal concept (A, B)

4 {4} {4} ({4}, {2, 4})
3 {3} {3} ({3}, {2, 3})
4 {3, 4} {3, 4} ({3, 4}, {2})
2 {2} {1, 2} non-canonic generation
1 {1} {1} ({1}, {1, 3, 4})
4 {1, 4} {1, 4} ({1, 4}, {4})
3 {1, 3} {1, 2, 3} non-canonic generation
2 {1, 2} {1, 2} ({1, 2}, {1, 3})
4 {1, 2, 4} {1, 2, 3, 4} non-canonic generation
3 {1, 2, 3} {1, 2, 3} ({1, 2, 3}, {3})
4 {1, 2, 3, 4} {1, 2, 3, 4} ({1, 2, 3, 4}, {})

whose extents are lectically less are generated first). By definition A is lectically
less than B if A ⊆ B, or B 6⊂ A and min((A ∪ B) \ (B ∩ A)) ∈ A. Note that
the NextClosure algorithm computes concepts in a different lectic order: A is
lectically less than B if min((A ∪ B) \ (B ∩ A)) ∈ B. The order in which concepts
are generated by CbO is beneficial when the line diagram is constructed: the
first generation of the concept is always canonical, which makes it possible to
find a concept in the tree and to draw appropriate diagram edges. NextClo-
sure-like lectic order allows binary search, which is helpful when the diagram
graph has to be generated after the generation of all concepts.

Algorithm 2 Close by One

Input: K = (G, M, I) is a context
Output: L is the concept set
1: L := ∅
2: for all g ∈ G do
3: Process({g}, g, ({g}00 , g))
4: end for
5: return L

The time complexity of Close by One (CbO) is O(|G|2 |M ||L|), and its
polynomial delay is O(|G|3 |M |).
The generation protocol of CbO in a tree-like form is given in Fig. 6. Each
closed set of objects (extent) can be read from the tree by following the path
from the root to the corresponding node. Square bracket ] means that first prime
operator has been applied after addition of the lectically next object g to the
set A of the parent node and bracket ) shows which object have been added
after application of second prime operator, i.e. between ] and ) one can find
(A ∪ {g})00 \ (A ∪ {g}). A non-canonic generation can be identified by simply
checking whether there is an object between ] and ) that less than g w.r.t. <.
 Introduction to Formal Concept Analysis and Its Applications in IR 13

Algorithm 3 Process(A, g, (C, D)) with C = A00 and D = A0 and ¡ the lexical
order on object names
Input: K = (G, M, I) is a context
Output: L is the concept set C = A00 , D = A0
1: if {h|h ∈ C \ A ∧ g < h} = ∅ then
2: L := L ∪ {(C, D)}
3: end if
4: for all f ∈ {h|h ∈ G \ A ∧ g < h} do
5: Z := C ∪ {f }
6: Z := D ∩ {f }0
7: X := Y 0
8: Process(Z, f, (X, Y ))
9: end for

One can note that the traverse of the generation tree is done in a depth-first
search manner.

1]) 2]) 3]4) 4])

2]) 3]24) 4]) 3]4) 4]3)

3]4) 4]3)

Fig. 6. The tree of CbO protocol for the context of geometric figures. Non-canonic
generations are drawn in boxes.

After the inception of the first batch algorithms, the broadened FCA inven-
tory includes efficient incremental algorithms [21] and the distributed versions
of NextClosure and CbO for MapReduce [22,23].

3.2 Many-valued contexts and concept scaling

Definition 15. A many-valued context (G, M, W, I) consists of sets G, M
and W and a ternary relation I between those three sets, i.e. I ⊆ G × M × W ,
for which it holds that (g, m, w) ∈ I and (g, m, v) ∈ I always imply w = v The
fact (g, m, w) ∈ I means “the attribute m takes value w for object g”, simply
written as m(g) = w.

Definition 16. A (conceptual) scale for the attribute m of a many-valued

context is a (one-valued) context Sm = (Gm , Mm , Im ) with m(G) = {m(g)|∀g ∈
14 Dmitry I. Ignatov

Table 2. Many-valued context of university subjects

G / M Gender Age Subject Mark

1 M 19 Math 8
2 F 20 CS 9
3 F 19 Math 7
4 M 20 CS 10
5 F 21 Data Mining 9

G} ⊆ Gm . The objects of a scale are called scale values, the attributes are called
scale attributes.
Nominal scale is defined by the context (Wm , Wm , =).
This type of scaling is suitable for binary representation of nominal (cate-
gorical) attributes like color. For the context of university subjects, the subjects
can be scaled by nominal scaling as below.
Math
=
DM
CS

Math ×
CS ×
DM ×
A particular case of nominal scaling is the so called dichotomic scaling,
which is suitable for attributes with two mutually exclusive values like “yes” and
“no”. In our example, the attribute Gender can be scaled in this way.
M F
M ×
F ×

Ordinal scale is given by the context (Wm , Wm , ≤) where denotes classical

real number order. For our example, the attributes age and mark can be scaled
by this type of scale.

≤ 7 ≤ 8 ≤ 9 ≤ 10
≤ 21 ≤ 20 ≤ 19
7 × × × ×
19 × × ×
8 × × ×
20 × ×
9 × ×
21 ×
10 ×

Interordinal scale is given by (Wm , Wm , ≤)|(Wm , Wm , ≥) where | denotes

the apposition of two contexts.
This type of scale can be used as an alternative for ordinal scaling like in
example below.
 Introduction to Formal Concept Analysis and Its Applications in IR 15

≤7≤8≤9 ≤ 10 ≥7 ≥ 8 ≥ 9 ≥ 10
7 × × × × ×
8 × × × × ×
9 × × × × ×
10 × × × × ×

In some domains, e.g., in psychology or sociology there is similar biordinal

(bipolar) scaling, which is a good representation of attributes with so called
polarvalues “agree”, “rather agree”, “disagree”, and “rather disagree”.
There is a special type of scale, contranominal scale, which is rare case
in real data, but has important theoretical meaning. Its context is given by
inequality relation, i.e. ({1, . . . , n}, {1, . . . , n}, 6=), and the example for n = 4 is
given below.

6= 1 2 3 4
1 ×××
2 × ××
3 ×× ×
4 ×××

In fact, this type of contexts gives rise to 2n formal concepts and can be used
for testing purposes.
The resulting scaled (or plain) context for our university subjects example is
below. Note that the Mark attribute is scaled by interordinal scale.

M F ≤ 19 ≤ 20 ≤ 21 Math CS DM ≤ 7 ≤ 8 ≤9≤ 10 ≥7 ≥ 8 ≥ 9 ≥ 10
1 × × × × × × × × × ×
2 × × × × × × × × ×
3 × × × × × × × × × ×
4 × × × × × × × × ×
5 × × × × × × × ×

3.3 Attribute Dependencies

Definition 17. Implication A → B, where A, B ⊆ M holds in context (G, M, I)
if A0 ⊆ B 0 , i.e., each object having all attributes from A also has all attributes
from B.
Example 4. For the context of geometric figures one may check that the following
implication holds: abc → d, b → c, cd → b. Note that for brevity we have omitted
curly brackets around and commas between elements of a set attributes.
Exercise 6. Find three more implications for the context of geometric figures. 
Implications satisfy Armstrong rules or inference axioms [24,25]:
X→Y
(reflexivity), (augmentation),
X→X X ∪Z →Y
16 Dmitry I. Ignatov

X → Y, Y ∪ Z → W
(pseudotransitivity).
X ∪Z →W
An inference axiom is a rule that states if certain implications are valid in
the context, then certain other implications are valid.

Example 5. Let us check that the first and second Armstrong axioms fulfill for
implication over attributes.
Since X 0 ⊆ X 0 it is always true that X → X.
For the second rule we have X 0 ⊆ Y 0 . Applying property 4 from Proposition 1
we have: (X ∪ Z)0 = X 0 ∩ Z 0 . Since X 0 ∩ Z 0 ⊆ X 0 , we prove that X 0 ∩ Z 0 ⊆ Y 0 .
This implies X ∪ Z → Y .

Exercise 7. 1. Prove by applying Armstrong rules that A1 → B1 and A2 → B2

imply A1 ∪ A2 → B1 ∪ B2 . 2. Check the third axiom by using implication
definition. 

Definition 18. An implication cover is a subset of implications from which

all other implications can be derived by means of Armstrong rules.
An implication base is a minimal (by inclusion) implication cover.

Definition 19. A subset of attributes D ⊆ M is a generator of a closed subset

of attributes B ⊆ M , B 00 = B if D ⊆ B, D00 = B = B 00 .
A subset D ⊆ M is a minimal generator if for any E ⊂ D one has
E 00 6= D00 = B 00 .
Generator D ⊆ M is called nontrivial if D 6= D00 = B 00 .
Denote the set of all nontrivial minimal generators of B by nmingen(B).
Generator implication cover looks as follows:
{F → (F 00 \ F ) | F ⊆ M, F ∈ nmingen (F 00 )}.

Example 6. For the context of geometric figures one may check that b is a mini-
mal nontrivial generator for bc, The set ab is a minimal nontrivial generator for
abcd, but abc, abd, and acd are its nontrivial generators. 

Exercise 8. For the context of geometric figures find all minimal generators and
obtain its generator implication cover. 

Definition 20. The Duquenne-Guigues base is an implication base where

each implication is a pseudo-intent [26].
A subset of attributes P ⊆ M is called a pseudo-intent if P 6= P 00 and for
any pseudo-intent Q such that Q ⊂ P one has Q00 ⊂ P .
The Duquenne-Guigues base looks as follows:
{P → (P 00 \ P ) | P is a pseudo-intent }.

The Duquenne-Guigues base is a minimum (cardinality minimal) implication

base.
 Introduction to Formal Concept Analysis and Its Applications in IR 17

Table 3. Finding pseudo-itents for the context of geometric figures

B B0 B 00 B is pseudo-intent?
∅ 1234 ∅ No, it’s not.
a 12 a No, it’s not.
b 34 bc Yes, it is.
c 234 c No, it’s not.
d 14 d No, it’s not.
ab ∅ abcd No, it’s not.
ac 2 ac No, it’s not.
ad 1 ad No, it’s not.
bc 34 bc No, it’s not.
bd 4 bcd No, it’s not.
cd 4 bcd Yes, it is.
abc ∅ abcd Yes, it is.
abd ∅ abcd No, it’s not.
acd ∅ abcd No, it’s not.
bcd 4 bcd No, it’s not.
abcd ∅ abcd No, it’s not.

Example 7. Let us find all pseudo-intents for the context of geometric figures.
We build a table (Table 3) with B and B 0 0; it is clear that all closed sets are not
pseudo-intents by the definition. Since we have to check the containment of a
pseudo-intent in the generated pseudo-intents recursively, we should start with
the smallest possible set, i.e. ∅.
Thus, {b} is the first non-closed set in our table and the second part of pseudo-
intent definition fulfills trivially – there is no another pseudo-intent contained in
{b}. So, the whole set of pseudo-intents is {b, cd, abc}. 

Exercise 9. Write down the Duquenne-Guigues base for the context of geometric
figures. Using Armstrong rules and the obtained Duquenne-Guigues base, deduce
the rest implications of the original context. 

For recent efficient algorithm of finding the Duquenne-Guigues base see [27].

Implications and functional dependencies Data dependencies are one way to

reach two primary purposes of databases: to attenuate data redundancy and
enhance data reliability [25]. These dependencies are mainly used for data nor-
malisation, i.e. their proper decomposition into interrelated tables (relations).
The definition of functional dependency [25] in terms of FCA is as follows:

Definition 21. X → Y is a functional dependency in a complete many-

valued context (G, M, W, I) if the following holds for every pair of objects g, h ∈
G:
(∀m ∈ X m(g) = m(h)) ⇒ (∀n ∈ Y n(g) = n(h)).

Example 8. For the example given in Table 2 the following functional dependen-
cies hold: Age → Subject, Subject → Age, M ark → Gender.
18 Dmitry I. Ignatov

The first two functional dependencies may have sense since students of the
same year may study the same subjects. However, the last one says Gender is
functionally dependent by Mark and looks as a pure coincidence because of the
small dataset.
The reduction of functional dependencies to implications:

Proposition 3. For a many-valued context (G, M, W, I), one defines the context
KN := (P2 (G), M, IN ), where P2 (G) is the set of all pairs of different objects
from G and IN is defined by

{g, h}IN m :⇔ m(g) = m(h).

Then a set Y ⊆ M is functionally dependent on the set X ⊆ M if and only if

the implication X → Y holds in the context KN .

Example 9. Let us construct the context KN for the many-valued context of

geometric figures.
Subject
Gender

Mark
Age

{1,2}
{1,3} × ×
{1,4} ×
{1,5}
{2,3} ×
{2,4} × ×
{2,5} × ×
{3,4}
{3,5} ×
{4,5}

One may check that the following implications hold: Age → Subject, Subject →
Age, M ark → Gender, which are the functional dependencies that we so in ex-
ample 8. 

An inverse reduction is possible as well.

Proposition 4. For a context K = (G, M, I) one can construct a many-valued

context KW such that an implication X → Y holds if and only if Y is functionally
dependent on X in KW .

Example 10. To fulfill the reduction one may build the corresponding many-
valued context in the following way:
1. Replace all “×” by 0s. 2. In each row, replace empty cells by the row
number starting from 1. 3. Add a new row filled by 0s.
 Introduction to Formal Concept Analysis and Its Applications in IR 19

abcd
1 0110
2 0202
3 3003
4 4000
5 0000

Exercise 10. Check the functional dependencies from the previous example co-
incide with the implications of the context of geometric figures. 

More detailed tutorial on FCA and fuctional dependencies is given in [28].

4 FCA tools and practice

In this section, we provide a short summary of ready-to-use software that sup-

ports basic functionality of Formal Concept Analysis.

– Software for FCA: Concept Explorer, Lattice Miner, ToscanaJ, Galicia, FCART
etc.
– Exercises.

Concept Explorer. ConExp 5 is probably one of the most user-friendly FCA-

based tools with basic functionality; it was developed in Java by S. Yevtushenko
under Prof. T. Taran supervision in the beginning of 2000s [29]. Later on it has
been improved several times, especially from lattice drawing viewpoint [30].
Now the features the following functionality:

– Context editing (tab separated and csv formats of input files are supported
as well);
– Line diagrams drawing (allowing their import as image snapshots and even
text files with nodes position, edges and attributes names, but vector-based
formats are not supported);
– Finding the Duquenne-Guigues base of implications;
– Finding the base of association rules that are valid in a formal context;
– Performing attribute exploration.

It is important to note that the resulting diagram is not static and one
may perform exploratory analysis in an interactive manner selecting interesting
nodes, moving them etc. In Fig. 7, the line diagram of the concept lattice of
interordinal scale for attribute Mark drawn by ConExp is shown. See more details
in Fig. [31].
There is an attempt to reincarnate ConExp 6 by modern open software tools.
5
http://conexp.sourceforge.net/
6
https://github.com/fcatools/conexp-ng/wiki
20 Dmitry I. Ignatov

Fig. 7. The line diagram of the concept lattice for the interordinal scale of student
marks drawn by ConExp.

ToscanaJ. The ToscanaJ7 project is a result of collaboration between two groups

from the Technical University of Darmstadt and the University of Queensland,
which aim was declared as “to give the FCA community a platform to work
with” [32] and “the creation of a professional tool, coming out of a research
environment and still supporting research” [33].
This open project has a long history with several prototypes [34] and now it
is a part of an umbrella framework for conceptual knowledge processing, Tockit8 .
As a result, it is developed in Java, supports different types of database connec-
tion via JDBC-ODBC bridge and contains an embedded database engine [33].
Apart from ConExp, it features work with multi-valued contexts, conceptual
scaling, and nested line diagrams.
In Fig. 8 one can see the nested line diagram for two scales from the university
subjects multi-valued context, namely for two attributes, Gender and Subject.
Via PDF printing facilities it is possible to print out line diagrams in a vector
graphic form.

Galicia. Galicia9 was “intended as an integrated software platform including

components for the key operations on lattices that might be required in prac-
tical applications or in more theoretically-oriented studies”. Thus in addition
to basic functionality of ConExp, it features work with multi-valued contexts
7
http://toscanaj.sourceforge.net/
8
http://www.tockit.org/
9
http://www.iro.umontreal.ca/~galicia/
 Introduction to Formal Concept Analysis and Its Applications in IR 21

Math
CS DM

M F

1 3
4 2 5

Fig. 8. The nested line diagram for the two one-attribute subcontexts of the context
of university subjects. The outer diagram is for Gender attribute, and the inner one is
for Subject.

and conceptual scaling, iceberg lattices (well-known in Data Mining commu-

nity), Galois hierarchies and relational context families, which are popular in
software engineering [35]. The software is open and its implementation in Java
is cross-platform aimed at “adaptability, extensibility and reusability”.
It is possible to navigate through lattice diagrams in an interactive manner;
the resulting diagrams contain numbered nodes and this is different from the
traditional way of line diagrams drawing. Another Galicia’s unique feature is 3D
lattice drawing. The diagram of the university subjects context after nominal
scaling of all its attributes obtained in Galicia is depicted in Fig. 9. Galicia sup-
ports vector-based graphic formats, SVG and PDF. The authors of the program
paid substantial attention to algorithmic aspects and incorporated batch and in-
cremental algorithms into it. Various bases of implications and association rules
can be generated by the tool. Nested line diagrams are in the to do list.

Lattice Miner. This is another attempt to establish basic FCA functionality and
several specific features to the FCA community 10 [36].
The initial objective of the tool was “to focus on visualization mechanisms
for the representation of concept lattices, including nested line diagrams” 11 .
Thus, its interesting feature is multi-level nested line diagrams, which can help
to explore comparatively large lattices.
10
http://sourceforge.net/projects/lattice-miner/
11
https://en.wikipedia.org/wiki/Lattice_Miner
22 Dmitry I. Ignatov

0
¥ I={}
¥ 
¥ E={obj_0, obj_1, obj_2, obj_3, obj_4}
¥ 

2 4
1 3
¥ I={Gender=M}
¥  ¥ I={Age=20, Subject=CS}
¥ 
¥ I={Age=19, Subject=Math}
¥  ¥ I={Gender=F}
¥ 
¥ E={obj_0, obj_3}
¥  ¥ E={obj_1, obj_3}
¥ 
¥ E={obj_0, obj_2}
¥  ¥ E={obj_1, obj_2, obj_4}
¥ 

6
¥ I={Age=19, Gender=F, Mark=7, Subject=Math}
¥ 
5 8
¥ E={obj_2}
¥ 
¥ I={Age=19, Gender=M, Mark=8, Subject=Math} ¥ I={Gender=F, Mark=9}
¥ 
¥ 
¥ E={obj_0} ¥ E={obj_1, obj_4}
¥ 
¥ 
9
¥ I={Age=20, Gender=F, Mark=9, Subject=CS}
¥ 
7 ¥ E={obj_1}
¥ 
¥ I={Age=20, Gender=M, Mark=10, Subject=CS}
¥  11
¥ E={obj_3}
¥ 
¥ I={Age=21, Gender=F, Mark=9, Subject=DM}
¥ 
¥ E={obj_4}
¥ 

10

Fig. 9. The line diagram of concept lattice for the context of university subjects drawn
by Galicia.

After more than a decade of development, FCA-based software having differ-

ent features produced a lot of different formats thus requiring interoperability.
To this end, in analogy to Rosetta Stone, FcaStone 12 was proposed. It sup-
ports convertation between commonly used FCA file formats (cxt, cex, csc, slf,
bin.xml, and csx) and comma separated value (csv) files as well as converta-
tion concept lattices into graph formats (dot, gxl, gml, etc. for use by graph
editors such as yEd, jgraph, etc.) or into vector graphics formats (fig, svg, etc.
for use by vector graphics editors such as Xfig, Dia, Inkscape, etc.). It can also
be incorporated into a webpage script for generating lattices and line diagrams
online. Another example of a web-based ported system with basic functionality
including attribute exploration is OpenFCA13 .

FCART. Many different tools have been created and some of the projects are not
developing anymore but the software is still available; an interested reader can
refer Uta Priss’s webpage to find dozens of tools14 . However, new challenges such
as handling large heterogeneous datasets (large text collections, social networks
and media etc.) are coming and the community, which put a lot of efforts in
12
http://fcastone.sourceforge.net/
13
https://code.google.com/p/openfca/
14
http://www.fcahome.org.uk/fcasoftware.html
 Introduction to Formal Concept Analysis and Its Applications in IR 23

the development of truly cross-platform and open software, needs a new wave of
tools that adopts modern technologies and formats.
Inspired by the successful application of FCA-based technologies in text min-
ing for criminology domain [37], in the Laboratory for Intelligent Systems and
Structural Analysis, a tool named Formal Concept Analysis Research Toolbox
(FCART) is developing.
FCART follows a methodology from [38] to formalise iterative ontology-
driven data analysis process and to implement several basic principles:
1. Iterative process of data analysis using ontology-driven queries and interac-
tive artifacts such as concept lattice, clusters, etc.
2. Separation of processes of data querying (from various data sources), data
preprocessing (via local immutable snapshots), data analysis (in interactive
visualizers of immutable analytic artifacts), and results presentation (in a
report editor).
3. Three-level extendability: settings customisation for data access components,
query builders, solvers and visualizers; writing scripts or macros; developing
components (add-ins).
4. Explicit definition of analytic artifacts and their types, which enables in-
tegrity of session data and links artifacts for the end-users.
5. Availability of integrated performance estimation tools.
6. Integrated documentation for software tools and methods of data analysis.
Originally, it was yet another FCA-based “integrated environment for knowl-
edge and data engineers with a set of research tools based on Formal Concept
Analysis” [39,40] featuring in addition work with unstructured data (including
texts with various metadata) and Pattern Structures [41]. In its current dis-
tributed version, FCART consists of the following parts:
1. AuthServer for authentication and authorisation.
2. Intermediate Data Storage (IDS) for storage and preprocessing of big datasets.
3. Thick Client for interactive data processing and visualisation in integrated
graphical multi-document user interface.
4. Web-based solvers for implementing independent resource-intensive compu-
tations.
The workflow is shown in Fig. 10.
The main questions are the following: Whether the product has only tech-
nological advantages or it really has fruitful methodology? Can it become open
in addition to its extendability? Can it finally handle big volumes of heteroge-
neous data in a suitable way for an FCART analyst? The answers to these posed
questions seem to be forthcoming challenging steps.

CryptoLatt. This tool15 was developed to help students and researchers from
neighbouring domains (e.g., Data Mining) to recognise cryptomorphisms in lattice-
based problems, i.e. to realise that a particular problem in one domain is “isomor-
phic” to some other in terms of lattice theory [42]. Thus, one of the well-known
15
http://www.cs.unic.ac.cy/florent/software.htm
4

interaction between user and external data storages (except local files) goes
through the IDS.
AuthServer and Intermediate Data Storage. IDS is responsible for storing
24 Dmitry I. Ignatov
and preprocessing chunks of data from various data sources in common format
(JSON collections). In the current release IDS provides a Web-API and uses
AuthServer for OAuth 2.0 authorization. There are many data storages, which
contain petabytes of collected data. For analyzing such Big Data analyst cannot
use any of the tools mentioned above. FCART provides different scenarios for
working with local and external data storages. In the previous release of FCART
analyst can work with quite small external datasets because all data from external
storages convert into JSON files and saves into IDS. In the current release, we
have improved strategies of working with external data. Now analyst can choose
between loading all data from external data storage to the IDS and accessing ex-
ternal data by chunks using paging mechanism. FCART analyst should specify the
way of accessing external data at the property page of the data generator (EDQD).

External Data Source External Data Set

External
Data Query
Description
FCART IDS FCART
Import/Export Intermediate JSON Collection
Tools Data Storage
Snapshot Profile
FCART Client
Local Session Data Snapshot
Import/Export
Data Base (Multivalued Context)
Tools
Scaling Query +
Graph Generators
Analytic Artifacts

Pattern Structures Binary Context

Clusters

Other artifacts Concept Lattice

Fig. 1. Main FCA workflow of FCART

Web Solvers and Web-API implementation. All interaction between Client,
Web Solvers and IDS goes through
Fig. 10. the Web-API.
FCART Client construct http-request to
workflow
the web-service. The http-request to web-service constructed from two parts: pre-
fix part and command part. Prefix part contains domain name and local path (e.g.
http://zeus2.hse.ru/lds/). The command part describes operations LDS has to do
and represents some function of web-service API. Using described commands
FCART client can query data from external data storage to the IDS.
 Introduction to Formal Concept Analysis and Its Applications in IR 25

cryptomorphisms in the FCA community is established between a lattice and

a binary relation, also known as the basic theorem of FCA. Note that even a
particular formal context, its concept lattice and set of implications represent
the same information about the underlying dataset but in a different way.

Exercise 11. Practice with Concept Explorer:

1. Input the context of geometric figures, build its concept lattice diagram and
find the Duquenne-Guigues base. Check whether the obtained base coincide with
the base found before. Play with different layouts and other drawing options like
labeling or node size.
2. Find real datasets where objects are described by nominal attributes and
select about 10 objects and 10 attributes from it. Prepare the corresponding
context, build the lattice diagram and find its implication base. Try to interpret
found concepts and dependencies. 

Exercise 12. Practice with ToscanaJ:

1. Use Elba tool from the latest version of ToscanaJ for creating two scaled
contexts for any two attributes of the context of university subjects. Save the
contexts. Then upload them into ToscanaJ and draw their nested line diagram.
The result should be similar to Fig. 8. 

Exercise 13. Practice with Galicia:

1. Perform tasks from exercise 11. 2. Compose the context of university subjects.
Scale it via Algorithms→Multi-FCA→Interactive Multi-FCA and build the lat-
tice diargam. The result should be identical to Fig. 9. 

5 FCA in Data Mining and Machine Learning

– Frequent Itemset Mining and Association Rules: FCA did it even earlier [43,44]
– Multimodal clustering (biclustering and triclustering) [45,46,47]
– FCA in Classification: JSM-method, version spaces∗16 , and decision trees∗
[48]
– Pattern Structures for data with complex descriptions [49,50]
– FCA-based Boolean Matrix Factorisation [51]
– Educational Data Mining case study [52]
– Exercises with JSM-method in QuDA (Qualitative Data Analysis): solving
classification task [53]

5.1 Frequent Itemset Mining and Association Rules

Knowledge discovery in databases (KDD) is introduced as the non-trivial extrac-

tion of valid, implicit, potentially useful and ultimately understandable informa-
tion in large databases [54]. Data mining is a main step in KDD, and in its turn
association rules and frequent itemset mining are among the key techniques in
16
not covered here
26 Dmitry I. Ignatov

Data Mining. The original problem for association rules mining is market basket
analysis. In early 90s, since the current level of technologies made it possible to
store large amount of transactions of purchased items, companies started their
attempts to use these data to facilitate their typical business decisions concern-
ing “what to put on sale, how to design coupons, how to place merchandise
on shelves in order to maximize the profit”[55]. So, firstly this market basket
analysis problem was formalised in [55] as a task of finding frequently bought
items together in a form of rules “if a customer buys items A, (s)he also buys
items B”. One of the first and rather efficient algorithms of that period was
proposed in [43], namely Apriori. From the very beginning these rules are toler-
ant to some number of exceptions, they were not strict as implications in FCA.
However, several years before, in [44], Michael Luxenburger introduced partial
implications motivated by more general problem statement, “a generalisation of
the theory of implications between attributes to partial implications” since “in
data analysis the user is not only interested in (global) implications, but also in
“implications with a few exceptions””. The author proposed theoretical treat-
ment of the problem in terms of Formal Concept Analysis and was guided by
the idea of characterisation of “sets of partial implications which arise from real
data” and “a possibility of an “exploration” of partial implications by a com-
puter”. In addition, he proposed a minimal base of partial implications known
as Luxenburger’s base of association rules as well.

Definition 22. Let K := (G, M, I) be a context, where G is a set of objects, M

is a set of attributes (items), I ⊆ G × M An association rule of the context
K is an expression A → B, where A, B ⊆ M and (usually) A ∩ B = ∅.

Definition 23. (Relative) support of an association rule A → B defined as

|(A ∪ B)0 |
supp(A → B) = .
|G|

The value of supp(A → B) shows which part of G contains A ∪ B. Often

support can be given in %.

Definition 24. (Relative) confidence of an association rule A → B defined as

|(A ∪ B)0 |
conf (A → B) = .
|A0 |

This value conf (A → B) shows which part of objects that possess A also
contains A ∪ B. Often confidence can be given in %.

Example 11. An object-attribute table of transactions.

 Introduction to Formal Concept Analysis and Its Applications in IR 27

Cakes

Chips
Müsli
Milk
Beer
c1 × ×
c2 ×××
c3 × × × ×
c4 × × × ×
c5 ××× ×

– supp({Beer, Chips}) = 3/5

|({Cakes, Müsli}∪{Milk})0 | |{C2,C5}|
– supp({Cakes, Müsli } → { Milk }) = |G| = 5 =
2/5
|({Cakes, Müsli}∪{Milk})0 | |{c2 ,c5 }|
– conf ({Cakes, Müsli } → { Milk }) = = |{c =
|{Cakes, Müsli}0 | 2 ,c5 }|
1


The main task of association rules mining is formulated as follows: Find all
association rules of a context, where support and confidence of the rules are
greater than predefined thresholds, min-confidence and min-support, denoted as
min conf and min supp, respectively [55]
Proposition 5. (Association rules and implications)
Let K be a context, then its associations rules under condition min supp =
0% and min conf = 100% are implications of the same context.
c
Sometimes an association rule can be written as A −→ B, where c and s are
s
confidence and support of the given rule.
Two main steps of association rules mining are given below:

1. Find frequent sets of attributes (frequent itemsets), i.e. sets of attributes

(items) that have support greater than min supp.
2. Building association rules based on found frequent itemsets.

The first step is the most expensive, the second one is rather trivial.
The well-known algorithm for frequent itemset mining is Apriori [43] uses
the antimonotony property to ease frequent itemsets enumeration.
Property 1. (Antimonotony property) For ∀A, B ⊆ M and A ⊆ B ⇒ supp(B) ≤
supp(A).
This property implies the following facts:
– The larger set, the smaller support it has or its support remains the same;
– Support of any itemset is not greater than a minimal support of any its
subset;
– Aa itemset of size n is frequent if and only if all its (n − 1)-subsets are
frequent.
28 Dmitry I. Ignatov

The Apriori algorithm finds all frequent itemsets.

It is check iteratively the set of all itemsets in a levelwise manner. At each
iteration one level is considered, i.e. a subset of candidate itemsets Ci is com-
posed by collecting the frequent itemsets discovered during the previous iteration
(AprioriGen procedure). Then supports of all candidate itemsets are counted,
and the infrequent ones are discarded.

Algorithm 4 Apriori(Context, min supp)

Input: Context, min supp is a minimal support
Output: all frequent itemsets IF
1: C1 ← 1-itemsets
2: i ← 1
3: while Ci 6= ∅ do
4: SupportCount(Ci )
5: Fi ← {f ∈ Ci |f.support ≥ min supp}
6: {Fi is a set of frequent i-itemsets}
7: Ci+1 ← AprioriGen(Fi )
8: {Ci is a set of (i + 1)-candidates}
9: i++
10: end while
S
11: IF ← Fi
12: return IF

For frequent itemsets of size i, procedure AprioriGen finds (i + 1)-supersets

and returns only the set of potentially frequent candidates.

Algorithm 5 AprioriGen(Fi )
Input: Fi is a set of frequent i-itemsets
Output: Ci+1 is a set of (i + 1)-itemsets candidates
1: insert into Ci+1 {union}
2: select p[1], p[2], . . . , p[i], q[i]
3: from Fi .p, Fi .q
4: where p[1] = q[1], . . . , p[i − 1] = q[i − 1], p[i] < q[i]
5: for all c ∈ Ci+1 do
6: {elimination}
7: S ← (i − 1)-itemset c
8: for all s ∈ S do
9: if s 6∈ Fi then
10: Ci+1 ← Ci+1 \ c
11: end if
12: end for
13: end for
14: return Ci+1
 Introduction to Formal Concept Analysis and Its Applications in IR 29

Example 12. Union and elimination steps of AprioriGen for a certain context.

– The set of frequent 3-itemsets: F3 = {{a, b, c}, {a, b, d}, {a, c, d}, {a, c, e}, {b, c, d}}.
– The set of candidate 4-itemsets (union step): C4 = {{a, b, c, d}, {a, c, d, e}}.
– The remaining candidate is C4 = {{a, b, c, d}}, since is eliminated {a, c, d, e}
because {c, d, e} 6∈ F3 (elimination step).

The worst-case computational complexity of the Apriori algorithm is O(|G||M |2 2|M |

since all the itemsets may be frequent. However, it takes only O(|M |) datatable
scans compared to O(2| M |) for brute-force method.
Rules extraction is based on frequent itemsets.
Let F be a frequent 2-itemset. We compose a rule f → F \ f if

supp(F )
conf (f → F \ f ) = ≥ min conf, wheref ⊂ F.
supp(f )
supp(F )
Property 2. Confidence conf (f → F \ f ) = supp(f ) is minimal, when supp(f ) is
maximal.

– Confidence is maximal when rule consequent F \ f consists of one attribute

(1-itemset). The subsets of such an consequent have greater support and it
turn smaller confidence.
– Recursive procedure of rules extraction starts with (|F | − 1)-itemset f ful-
filling min conf and min sup; then, it forms the rule f → F \ f and checks
all its subsets (|F | − 2)-itemset (if any) and so on.

Exercise 14. Find all frequent itemsets for the customers context with with Apri-
ori algorithm and min sup = 1/3. 

Condensed representation of frequent itemsets According to basic results from

Formal Concept Analysis, it is not necessary count the support of all frequent
itemsets. Thus, it is possible to derive from some known supports the supports of
all other itemsets: it is enough to know the support of all frequent concept intents.
And it is also not necessary to compute all frequent itemsets for solving the
association rule problem: it is sufficient to consider the frequent concept intents
that also called closed itemsets in Data Mining. In fact, closed itemsets was
independently discovered by three groups of researches in the late 90s [56,57,58].
Let K = (G, M, I) be a formal context.

Definition 25. A set of attributes F C ⊆ M is called frequent closed item-

set, if supp(F C) ≥ min supp and there is no any F such that F ⊃ F C and
supp(F ) = supp(F C).

Definition 26. A set of attributes M F C ⊆ M is called maximal frequent

itemset if it is frequent and there is no any F such thatF ⊃ F M C and supp(F ) ≥
min supp.
30 Dmitry I. Ignatov

Proposition 6. In a formal context K, MFC ⊆ F C ⊆ F, where MFC is the

set of maximal frequent itemset, FC is the set of frequent closed itemsets, and
F is the set of frequent itemsets of K with a minimal support min supp.

Proposition 7. The concept lattice of a formal context K is (isomorphic to) its

lattice of frequent closed itemsets with min supp = 0.

One may check that the lattices, whose diagrams are depicted in Fig. 11, are
isomorphic.

abcde(0)
2 1

acde(1) abce bce(1)

7 9 (1) 6 3
e 12
c

cde
(2)
ace(2) bce(2) bcd(2) a d b
10 5 8 14
c1 11 4 13 15

cd
(3)
ae(3) ce(3) bc(3)
11 4 13 15 10 5 8 14
c2

e(4) c(4)
3 12 7 9 6
c3 c4 c5

1
∅(5) 2

Fig. 11. The line diagrams of the lattice of closed itemsets (left, with support given
in parentheses) and the concept lattice for the customers context (right, with reduced
labeling)

The set of all frequent concepts of the context K for the threshold min sup is
also known as the “iceberg concept lattice” [59], mathematically it corresponds
to the order filter of the concept lattice. However, the idea of usage the size
of concept’s extent, intent (or even their different combinations) as a concept
quality measure is not new in FCA [60].
Of course, the application domain is not restricted to market basket analysis;
thus, the line diagram built in ConExp shows 25 largest concepts of visitors of
HSE website in terms of news websites in 2006.
For real datasets, association rules mining usually results in the large number
of rules. However, not all rules are necessary to present the information. Similar
compact representation can be used here; thus, one can represent all valid asso-
ciation rules by their subsets that called bases. For example, the Luxenburger
base is a set of association rules in the form

{B1 → B2 |(B10 , B1 ) is an upper neighbour of concept (B20 , B2 )}.

 Introduction to Formal Concept Analysis and Its Applications in IR 31

Fig. 12. The line diagram of 25 largest concepts for the context of HSE web users

The rest rules and their support and confidence can be derived by some
calculus, which is not usually clear from the base definition.

Exercise 15. 1. Find the Luxenburger base for the customers context with min sup =
1/3 and min conf = 1/2. 2. Check whether Concept Explorer generates the as-
sociation rule base (for the same context) that consists of the Duquenne-Guigues
base and the Luxenburger base. 

One of the first algorithms that were explicitly designed to compute frequent
closed itemsets is Close [56]. Inspired by Apriori it traverses the database in
a level-wise manner and generates requent closed itemsets by computing the
closures of all minimal generators. See more detailed survey in [61].
It is interesting that after the years of co-existence, one of the Apriori’s
authors has started to apply FCA in text mining [62]. FCA is also included into
textbooks on data mining, see Chapter 8&9 in [63].
Another interesting subdomain of frequent pattern mining, where lattice-
based methods are successfully used, is so called sequential pattern mining
[64,41].

Multimodal clustering (biclustering and triclustering) Clustering is an

activity for finding homogeneous groups of instances in data. In machine learn-
ing, clustering is a part of so called unsupervised learning. The widely adopted
idea of cluster relates to instances in a feature space. A cluster in this space is
a subset of data instances (points) that are relatively close to each other but
relatively far from other data points. Such feature space clustering algorithms
are a popular tool in marketing research, bioinformatics, finance, image analy-
sis, web mining, etc. With the growing popularity of recent data sources such as
32 Dmitry I. Ignatov

biomolecular techniques and Internet, other than instance-to-feature data appear

for analysis.
One example is gene expression matrices, entries of which show expression
levels of gene material captured in a polymerase reaction. Another example
would be n-ary relations among several sets of entities such as:
– Folksonomy data [65] capturing a ternary relation among three sets: users,
tags, and resources;
– Movies database IMDb (17 ) describing a binary relation of “relevance” be-
tween a set of movies and a set keywords or a ternary relation between sets
of movies, keywords and genres;
– product review websites featuring at least three itemsets (product, product
features, product-competitor);
– job banks comprising at least four sets (jobs, job descriptions, job seekers,
seeker skills).
For two-mode case other cluster approaches demonstrates growing popularity.
Thus the notion of bicluster in a data matrix (coined by B. Mirkin in [66], p.
296) represents a relation between two itemsets. Rather than a single subset of
entities, a bicluster features two subsets of different entities.
In general, the larger the values in the submatrix, the higher interconnection
between the subsets, the more relevant is the corresponding bicluster. In the
relational data, presence-absence facts represented by binary 1/0 values and this
condition expresses the proportion of unities in the submatrix, its “density”:
the larger, the better. It is interesting that a bicluster of the density 1 is a
formal concept if its constituent subsets cannot be increased without a drop
in the density value, i.e. a maximal rectangle of 1s in the input matrix w.r.t.
permutations of its rows and columns [3]. Usually one of the related sets of
entities is a set of objects, the other one is a set of attributes. So, in contrary to
ordinary clustering, bicluster (A, B) captures similarity (homogeneity) of objects
from A expressed in terms of their common (or having close values) attributes
B, which usually embrace only a subset of the whole attribute space.
Obviously, biclusters form a set of homogeneous chunks in the data so that
further learning can be organized within them. The biclustering techniques and
FCA machinery are being developed independently in independent communi-
ties using different mathematical frameworks. Specifically, the mainstream in
Formal Concept Analysis is based on ordered structures, whereas biclustering
relies on conventional optimisation approaches, probabilistic and matrix algebra
frameworks [67,68]. However, in fact these different frameworks considerably
overlap in applications, for example: finding co-regulated genes over gene ex-
pression data [67,69,70,71,72,73,68], prediction of biological activity of chemical
compounds [74,75,76,77], text summarisation and classification [78,79,80,81,82],
structuring websearch results and browsing navigation in Information Retrieval
[15,83,84,4], finding communities in two-mode networks in Social Network Anal-
ysis [85,86,87,88,89] and Recommender Systems [90,91,92,93,94].
17
www.imdb.com
 Introduction to Formal Concept Analysis and Its Applications in IR 33

For example, consider a bicluster definition from paper [70]. Bi-Max algo-
rithm described in [70] constructs inclusion-maximal biclusters defined as
follows:

Definition 27. Given m genes, n situations and a binary table e such that
eij = 1 (gene i is active in situation j) or eij = 0 (gene i is not active in situation
j) for all i ∈ [1, m] and j ∈ [1, n], the pair (G, C) ∈ 2{1,...,n} × 2{1,...,m} is called
an inclusion-maximal bicluster if and only if (1) ∀i ∈ G, j ∈ C : eij = 1 and
(2) @(G1 , C1 ) ∈ 2{1,...,n} × 2{1,...,m} with (a) ∀i1 ∈ G1 , ∀j1 ∈ C1 : ei1 j1 = 1 and
(b) G ⊆ G1 ∧ C ⊆ C1 ∧ (G1 , C1 ) 6= (G, C).

Let us denote by H the set of genes (objects in general), by S the set of

situations (attributes in general), and by E ⊆ H × S the binary relation given
by the binary table e, |H| = m, |S| = n. Then one has the following proposition:

Proposition 8. For every pair (G, C), G ⊆ H, C ⊆ S the following two state-
ments are equivalent.
1. (G, C) is an inclusion-maximal bicluster of the table e;
2. (G, C) is a formal concept of the context (H, S, E).

Exercise 16. Prove Proposition 8.

Object-Attribute-biclustering Another example is OA-biclustering proposed

in [95,96] as a reliable relaxation of formal concept.

Definition 28. If (g, m) ∈ I, then (m0 , g 0 ) is called an object-attribute or OA-

0
×g 0 )|
bicluster with density ρ(m0 , g 0 ) = |I∩(m
|m0 |·|g 0 | .

Here are some basic properties of OA-biclusters.

Proposition 9. 1. 0 ≤ ρ ≤ 1.
2. OA-bicluster (m0 , g 0 ) is a formal concept iff ρ = 1.
3. if (m0 , g 0 ) is a OA-bicluster, then (g 00 , g 0 ) ≤ (m0 , m00 ).

Exercise 17. a. Check that properties 1. and 2. from Proposition 9 follow directly
by definitions. b. Use antimonotonicity of (·)0 to prove 3. 

In figure 13 one can see the structure of the OA-bicluster for a particular pair
(g, m) ∈ I of a certain context (G, M, I). In general, only the regions (g 00 , g 0 ) and
(m0 , m00 ) are full of non-empty pairs, i.e. have maximal density ρ = 1, since they
are object and attribute formal concepts respectively. Several black cells indicate
non-empty pairs which one may found in such a bicluster. It is quite clear, the
density parameter ρ would be a bicluster quality measure which shows how many
non-empty pairs the bicluster contains.

Definition 29. Let (A, B) ∈ 2G × 2M be an OA-bicluster and ρmin be a non-

negative real number, such that 0 ≤ ρmin ≤ 1, then (A, B) is called dense if it
satisfies the constraint ρ(A, B) ≥ ρmin .
34 Dmitry I. Ignatov

g'

m' g g''

m''

Fig. 13. Bicluster based on object and attribute closures

Order relation v on OA-biclusters is defined component-wise: (A, B) v

(C, D) iff A ⊆ C and B ⊆ D.
Monotonicity (antimonotonicity) of constraints is often used in mining asso-
ciation rules for effective algorithmic solutions.

Proposition 10. The constraint ρ(A, B) ≥ ρmin is neither monotonic nor anti-
monotonic w.r.t. v relation.

Exercise 18. 1. To prove Proposition 10 for context K consider OA-biclusters

b1 = ({g1 , g3 , g4 , g5 }, {m1 , m4 , m5 }), b3 = (G, {m1 , m2 , m3 }) and b2 = (G, M ).
m1
m2
m3
m4
m5

g1 ×××××
g2 ××××
g3 × ××
g4 × ××
g5 × ××

2. Find generating pairs (g, m) for all these three biclusters. 

However, the constraint on ρmin has other useful properties.

If ρ = 0, this means that we consider the set of all OA-biclusters of the
context K. For ρmin = 0 every formal concept is “contained” in a OA-bicluster
of the context K, i.e., the following proposition holds.

Proposition 11. For each (Ac , Bc ) ∈ B(G, M, I) there exists a OA-bicluster

(Ab , Bb ) ∈ B such that (Ac , Bc ) v (Ab , Bb ).

Proof. Let g ∈ Ac , then by antimonotonicity of (·)0 we obtain g 0 ⊇ Bc .

Similarly, for m ∈ Bc we have m0 ⊇ Ac . Hence, (Ab , Bb ) v (m0 , g 0 ). 
 Introduction to Formal Concept Analysis and Its Applications in IR 35

The number of OA-biclusters of a context can be much less than the number
of formal concepts (which may be exponential in |G| + |M |), as stated by the
following proposition.

Proposition 12. For a given formal context K = (G, M, I) and ρmin = 0 the
largest number of OA-biclusters is equal to |I|, all OA-biclusters can be generated
in time O(|I| · (|G| + |M |)).

Proposition 13. For a given formal context K = (G, M, I) and ρmin > 0 the
largest number of OA-biclusters is equal to |I|, all OA-biclusters can be generated
in time O(|I| · |G| · |M |).

Algorithm 6 OA-bicluster computation

Input: K = (G, M, I) is a formal context, ρmin is a threshold density value of bicluster
density
Output: B = {(Ak , Bk )|(Ak , Bk ) is a bicluster}
1: B ← ∅
2: if ρmin = 0 then
3: for all (m, g) ∈ I do
4: B.Add(m0 , g 0 )
5: end for
6: else
7: for all (m, g) ∈ I do
8: if ρ(m0 , g 0 ) ≥ ρmin then
9: B.Add(m0 , g 0 )
10: end if
11: end for
12: end if
13: B.RemoveDuplicates()
14: return B

Algorithm 6 is a rather straightforward implementation by definition, which

takes initial formal context and minimal density threshold as parameters and
computes biclusters for each (object, attribute) pair in relation I. However, in its
latest implementations we effectively use hashing for duplicates elimination. In
our experiments on web advertising data, the algorithm produces 100 times less
patterns than the number of formal concepts. In general, for the worst case these
values are 2min(|G|,|M |) vs |I|. The time complexity of our algorithm is polinomial
(O(|I||G||M |)) vs exponential in the worst case for Bi-Max (O(|I||G||L| log |L|))
or CbO (O(|G|2 |M ||L|)), where |L| is a number of generated concepts which is
exponential in the worst case (|L| = 2min(|G|,|M |) ).

Triadic FCA and triclustering As we have mentioned, there are such data
sources as folksonomies, for example, a bookmarking website for scientific lit-
36 Dmitry I. Ignatov

erature Bibsonomy 18 [97]; the underlying structure includes triples (user, tag,
bookmark) like one in Fig. 14.

Fig. 14. An example of Bibsonomy relation for three paper, five authors and five tags.

Therefore, it can be useful to extend the biclustering and Formal Concept

Analysis to process relations among more than two datasets. A few attempts in
this direction have been published in the literature. For example, Zaki et al. [98]
proposed Tricluster algorithm for mining biclusters extended by time dimension
to real-valued gene expression data. A triclustering method was designed in [99]
to mine gene expression data using black-box functions and parameters coming
from the domain. In the Formal Concept Analysis framework, theoretic papers
[100,101] introduced the so-called Triadic Formal Concept Analysis. In [102], tri-
adic formal concepts apply to analyse small datasets in a psychological domain.
Paper [45] proposed rather scalable method Trias for mining frequent tricon-
cepts in Folksonomies. Simultaneously, a less efficient method on mining closed
cubes in ternary relations was proposed by Ji et al. [103]. There are several recent
efficient algorithms for mining closed ternary sets (triconcepts) and even more
general algorithms than Trias. Thus, Data-Peeler [104] is able to mine n-ary
formal concepts and its descendant mines fault-tolerant n-sets [105]; the latter
was compared with DCE algorithm for fault-tolerant n-sets mining from [106].
The paper [107] generalises n-ary relation mining to multi-relational setting in
databases using the notion of algebraic closure.
In triadic setting, in addition to set of objects, G, and set of attributes, M ,
we have B, a set of conditions. Let K = (G, M, B, I) be a triadic context,
where G, M , and B are sets, and I is a ternary relation: I ⊆ G × M × B. The
18
bibsonomy.org
 Introduction to Formal Concept Analysis and Its Applications in IR 37

triadic concepts of an triadic context (G, M, B, Y ⊆ G × M × B) are exactly

the maximal 3-tuples (A1 , A2 , A3 ) in 2G × 2M × 2B with A1 × A2 × A3 ⊆ Y with
respect to component-wise set inclusion [100,101]. The notion of n-adic concepts
can be introduced in the similar way to the triadic case [108].

Example 13. For the bibsonomy example, one of the triadic concepts is

({P oelmans, Elzinga}, {Domestic V iolence}, {paper3})

(see dotted edges on the graph in Fig.14). It means that both users Poelmans
and Elzinga marked paper 3 by the tag “Domestic Violence”. 

Guided by the idea of finding scalable and noise-tolerant triconcepts, we had

a look at triclustering paradigm in general for a triadic binary data, i.e. for
tricontexts as input datasets.
Suppose X, Y , and Z are some subsets of G, M , and B respectively.

Definition 30. Suppose K = (G, M, B, I) is a triadic context and Z ⊆ G, Y ⊆

M , Z ⊆ B. A triple T = (X, Y, Z) is called an OAC-tricluster. Traditionally,
its components are called (tricluster) extent, (tricluster) intent, and (tricluster)
modus, respectively.

The density of a tricluster T = (X, Y, Z) is defined as the fraction of all

triples of I in X × Y × Z:

|I ∩ (X × Y × Z)|
ρ(T ) := .
|X||Y ||Z|
Definition 31. The tricluster T is called dense iff its density is not less than
some predefined threshold, i.e. ρ(T ) ≥ ρmin .

The collection of all triclusters for a given tricontext K is denoted by T .

Since we deal with all possible cuboids in Cartesian product G × M × B, it
is evident that the number of all OAC-triclusters, |T |, is equal to 2|G|·|M |·|B| .
However not all of them are supposed to be dense, especially for real data which
are frequently quite sparse. Thus we have proposed two possible OAC-tricluster
definitions, which give us an efficient way to find within polynomial time a num-
ber of (dense) triclusters not greater than the number |I| of triples in the initial
data.
In [109], we have compared a set of triclustering techniques proposed within
Formal Concept Analysis and/or bicluster analysis perspectives: OAC-box [46],
Tribox [110], SpecTric [47] and a recent OAC-prime algorithm. This novel
algorithm, OAC-prime, overcomes computational and substantive drawbacks
of the earlier formal-concept-like algorithms. In our spectral approach (SpecTric
algorithm) we rely on an extension of the well-known reformulation of a bipartite
graph partitioning problem to the spectral partitioning of a graph (see, e.g. [78]).
For comparison purposes, we have proposed new developments in the following
components of the experiment setting:
38 Dmitry I. Ignatov

1. Evaluation criteria: The average density, the coverage, the diversity and the
number of triclusters, and the computation time and noise tolerance for the
algorithms.
2. Benchmark datasets: We use triadic datasets from publicly available internet
data as well as synthetic datasets with various noise models.

A preceding work was done in [111].

As a result we have not defined an absolute winning methods, but the multi-
criteria choice allows an expert to decide which of the criteria are most important
in a specific case and make a choice. Thus our experiments show that our Tri-
box and OAC-prime algorithms can be reasonable alternatives to triadic formal
concepts and lead to Pareto-effective solutions. In fact TriBox is better with
respect to noise-tolerance and the number of clusters, OAC-prime is the best
on scalability to large real-world datasets. In paper [112], an efficient version of
online OAC-prime has been proposed.
In our experiments we have used a context of top 250 popular movies from
www.imdb.com, objects are movie titles, attributes are tags, whereas conditions
are genres. Prime OAC-triclustering showed rather good results being one the
fastest algorithm under comparison.

Example 14. Examples of Prime OAC triclusters with their density indication
for the IMDB context are given below:

1. 36%, {The Shawshank Redemption (1994), Cool Hand Luke (1967), Ameri-
can History X (1998), A Clockwork Orange (1971), The Green Mile (1999)},
{Prison, Murder, Friend, Shawshank, Banker}, {Crime, Drama}
2. 56, 67%, {The Godfather: Part II (1974), The Usual Suspects (1995)}, {Cuba,
New York, Business, 1920s, 1950s}, {Crime, Drama, Thriller}
3. 60%, {Toy Story (1995), Toy Story 2 (1999)}, {Jealousy, Toy, Spaceman,
Little Boy, Fight}, {Fantasy, Comedy, Animation, Family, Adventure}

5.2 FCA in Classification

It is a matter of fact that Formal Concept Analysis helped to algebraically re-

think several models and methods in Machine Learning such as version spaces [113],
learning from positive and negative examples [74,48], and decision trees [48]. It
was also shown that concept lattice is a perfect search space for learning globally
optimal decision trees [114]. Already in early 90s both supervised and unsuper-
vised machine learning techniques and applications based on Formal Concept
Analysis were introduced in the machine learning community. E.g., in ML-related
venues there were reported results on the concept lattice based clustering in GA-
LOIS system that suited for information retrieval via browsing [115,116]. [117]
performed a comparison of seven FCA-based classification algorithms. [118] and
[119] propose independently to use FCA to design a neural network architec-
ture. In [120,121] FCA was used as a data preprocessing technique to transform
 Introduction to Formal Concept Analysis and Its Applications in IR 39

the attribute space to improve the results of decision tree induction. Note that
FCA helps to perform feature selection via conceptual scaling and has quite ev-
ident relations with Rough Sets theory, a popular tool for feature selection in
classification [122]. [123] proposed Navigala, a navigation-based approach for su-
pervised classification, and applied it to noisy symbol recognition. Lattice-based
approaches were also successfully used for classification of data with complex
descriptions such as graphs or trees [75,124]. Moreover, in [125] (Chapter 4,
“Concept Learning”) FCA is suggested as an alternative learning framework.

JSM-method of hypothesis generation The JSM-method proposed by Vik-

tor K. Finn in late 1970s was proposed as attempt to describe induction in purely
deductive form and thus to give at least partial justification of induction [126].
The method is named to pay respect to the English philosopher John Stuart
Mill, who proposed several schemes of inductive reasoning in the 19th century.
For example, his Method of Agreement, is formulated as follows: “If two or more
instances of the phenomenon under investigation have only one circumstance in
common, ... [it] is the cause (or effect) of the given phenomenon.”
The method proved its ability to enable learning from positive and negative
examples in various domains [127], e.g., in life sciences [74].
For RuSSIR audience, the example of the JSM-method application in paleog-
raphy might be especially interesting [128]: JSM was used for dating birch-bark
documents of 10–16 centuries of the Novgorod republic. There were five types of
attributes: individual letter features, features common to several letters, hand-
writing, language features (morphology, syntax, and typical errors), style (letter
format, addressing formulas and their key words).
Even though, the JSM-method was formulated in a mathematical logic set-
ting, later on the equivalence between JSM-hypotheses and formal concepts was
recognized [60].
The following definition of a hypothesis (“no counterexample-hypothesis”) in
FCA terms was given in [129].
Let K = (G, M, I) be a context. There are a target attribute w ∈ / M,

– positive examples, i.e. set G+ ⊆ G of objects known to have w,

– negative examples, i.e. set G− ⊆ G of objects known not to have w,
– undetermined examples, i.e. set Gτ ⊆ G of objects for which it is un-
known whether they have the target attribute or do not have it.

There are three subcontexts of K = (G, M, I), the first two are used for the
training sample: Kε := (Gε , M, Iε ), ε ∈ {−, +, τ } with respective derivation
operators (·)+ , (·)− , and (·)τ .

Definition 32. A positive hypothesis H ⊆ M is an intent of K+ not con-

tained in the intent g − of any negative example g ∈ G− : ∀g ∈ G− H 6⊆ g − .
Equivalently,
H ++ = H, H 0 ⊆ G+ ∪ Gτ .
40 Dmitry I. Ignatov

Negative hypotheses are defined similarly. An intent of K+ that is contained

in the intent of a negative example is called a falsified (+)-generalisation.

Example 15. In Table 4, there is a many-valued context representing credit scor-

ing data.
G+ = {1, 2, 3, 4}, G− = {5, 6, 7}, and Gτ = {8, 9, 10}. The target attribute
takes values + and − meaning “low risk” and “high risk” client, respectively.

Table 4. Many-valued classification context for credit scoring

G / M Gender Age Education Salary Target

1 M young higher high +
2 F middle special high +
3 F middle higher average +
4 M old higher high +
5 M young higher low −
6 F middle secondary average −
7 F old special average −
8 F young special high τ
9 F old higher average τ
10 M middle special average τ

To apply JSM-method in FCA terms we need to scale the given data. One
may use nominal scaling as below.

M F Y M i O HE Sp Se HS A L w w̄
g1 × × × × ×
g2 × × × × ×
g3 × × × × ×
g4 × × × × ×
g5 × × × × ×
g6 × × × × ×
g7 × × × × ×

Then we need to find positive and negative non-falsified hypotheses. If Fig. 15

there are two lattices of positive and negative examples for the input context,
respectively.
Shaded nodes correspond to maximal non-falsified hypotheses, i.e. they have
no upper neighbors being non-falsified hypotheses.
For K+ hypothesis {HE} is falsified since object g5 provides a counterexam-
ple, i.e. {HE} ⊆ g5− = {M, Y, HE, L}.
For K− hypothesis {A, F } is falsified since there is a positive counterexample,
namely {A, F } ⊆ g3+ = {F, M, HE, A}.

Undetermined examples gτ from Gτ are classified as follows:

 Introduction to Formal Concept Analysis and Its Applications in IR 41

HE HS
A,F
Mi,F M
L,HE,Y,M
O SE,Mi,Sp
g5
A Sp O Y g7 g6

g3 g2 g4 g1 HS

L,SE

Fig. 15. The line diagrams of the lattice of positive hypotheses (left) and the lattice
of negative hypotheses (right).

– If gττ contains a positive, but no negative hypothesis, then gτ is classified

positively (presence of target attribute w predicted).
– If gττ contains a negative, but no positive hypothesis, then gτ classified
negatively (absence of target attribute w predicted).
– If gττ contains both negative and positive hypotheses, or if gττ does not con-
tain any hypothesis, then object classification is contradictory or unde-
termined, respectively.

It is clear, for performing classification it is enough to have only minimal

hypotheses (w.r.t. ⊆), negative and positive ones.

Exercise 19. For the credit scoring context, classify all undetermined examples.


There is a strong connection between hypotheses and implications.

Proposition 14. A positive hypothesis h corresponds to an implication h →

{w} in the context K+ = (G+ , M ∪ {w}, I+ ∪ G+ × {w}).
A negative hypothesis h corresponds to an implication h → {w̄} in the context
K− = (G− , M ∪ {w̄}, I− ∪ G− × {w̄}).
Hypotheses are implications which premises are closed (in K+ or in K− ).

A detailed yet retrospective survey on JSM-method (in FCA-based and orig-

inal formulation) and its applications can be found in [14]. A further extension of
JSM-method to triadic data with target attribute in FCA-based formulation can
be found in [130,131]; there, the triadic extension of JSM-method used CbO-like
algorithm for classification in Bibsonomy data.
However, we saw that original data often need scaling, but, for example,
it is not evident what to do in case of learning with labeled graphs. To name
42 Dmitry I. Ignatov

a few problems of this kind we would mention structure-activity relationship

problems for chemicals given by molecular graphs and learning semantics from
graph-based (XML, syntactic tree) text representations. Motivated by search of
possible extensions of original FCA machinery to analyse data with complex
structure, Ganter and Kuznetsov proposed so called Pattern Structures [132].

5.3 Pattern Structures for data with complex descriptions

The basic definitions of Pattern Structures were proposed in [132].

Let G be a set of objects and D be a set of all possible object descriptions.
Let u be a similarity operator. It helps to work with objects that have non-
binary attributes like in traditional FCA setting, but those that have complex
descriptions like intervals [73], sequences [133] or (molecular) graphs [75]. Then
(D, u) is a meet-semi-lattice of object descriptions. Mapping δ : G → D assigns
an object g the description d ∈ (D, u).
A triple (G, (D, u), δ) is a pattern structure. Two operators (·) define Galois
connection between (2G , ⊆) and (D, u):

l
A = δ(g) for A ⊆ G (1)
g∈A

d = {g ∈ G|d v δ(g)} for d ∈ (D, u), where (2)

d v δ(g) ⇐⇒ d u δ(g) = d.

For a set of objects A operator 1 returns the common description (pattern)

of all objects from A. For a description d operator 2 returns the set of all objects
that contain d.
A pair (A, d) such that A ⊆ G and d ∈ (D, u) is called a pattern concept
of the pattern structure (G, (D, u), δ) iff A = d and d = A. In this case A
is called a pattern extent and d is called a pattern intent of a pattern concept
(A, d). Pattern concepts are partially ordered by (A1 , d1 ) ≤ (A2 , d2 ) ⇐⇒ A1 ⊆
A2 ( ⇐⇒ d2 v d1 ). The set of all pattern concepts forms a complete lattice
called a pattern concept lattice.

Intervals as patterns. It is obvious that similarity operator on intervals should

fulfill the following condition: two intervals should belong to an interval that
contains them. Let this new interval be minimal one that contains two original
intervals. Let [a1 , b1 ] and [a2 , b2 ] be two intervals such that a1 , b1 , a2 , b2 ∈ R,
a1 ≤ b1 and a2 ≤ b2 , then their similarity is defined as follows:

[a1 , b1 ] u [a2 , b2 ] = [min(a1 , a2 ), max(b1 , b2 )].

 Introduction to Formal Concept Analysis and Its Applications in IR 43

Therefore

[a1 , b1 ] v [a2 , b2 ] ⇐⇒ [a1 , b1 ] u [a2 , b2 ] = [a1 , b1 ]

 
⇐⇒ min(a1 , a2 ), max(b1 , b2 ) = [a1 , b1 ]
⇐⇒ a1 ≤ a2 and b1 ≥ b2 ⇐⇒ [a1 , b1 ] ⊇ [a2 , b2 ]

Note that a ∈ R can be represented by [a, a].

Interval vectors as patterns. Let us call p-adic vectors of intervals as interval vec-
tors. In this case for two interval vectors of the same dimension e = h[ai , bi ]ii∈[1,p]
and f = h[ci , di ]ii∈[1,p] we define similarity operation via the intersection of the
corresponding components of interval vectors, i.e.:

e u f = h[ai , bi ]ii∈[1,p] u h[ci , di ]ii∈[1,p] ⇐⇒ e u f = h[ai , bi ] u [ci , di ]ii∈[1,p]

Note that interval vectors are also partially ordered:

e v f ⇐⇒ h[ai , bi ]ii∈[1,p] v h[ci , di ]ii∈[1,p] ⇐⇒ [ai , bi ] v [ci , di ]

for all i ∈ [1, p].

Example 16. Consider as an example the following table of movie ratings:

Table 5. Movie rates

The Artist Ghost Casablanca Mamma Mia! Dogma Die Hard Leon
User1 4 4 5 0 0 0 0
User2 5 5 3 4 3 0 0
User3 0 0 0 4 4 0 0
User4 0 0 0 5 4 5 3
User5 0 0 0 0 0 5 5
User6 0 0 0 0 0 4 4

Each user of this table can be described by vector of ratings’ intervals. For
example, δ(u1 ) = h[4, 4], [4, 4], [5, 5], [0, 0], [0, 0], [0, 0], [0, 0]i. If some new user
u likes movie Leon, a movie recommender system would reply who else like
this movie by applying operator 2: [4, 5] Leon = {u5 , u6 }. Moreover, the sys-
tem would retrieve the movies that users 5 and 6 liked, hypothesizing that
they have similar tastes with u. Thus, operator 1 results in d = {u5 , u6 } =
h[0, 0], [0, 0], [0, 0], [0, 0], [0, 0], [4, 5], [4, 5]i, suggesting that Die Hard is worth watch-
ing for the target user u.
44 Dmitry I. Ignatov

Obviously, the pattern concept ({u5 , u6 }, d) describes a small group of like-

minded users and their shared preferences are stored in the vector d (cf. biclus-
ter). 
Taking into account constant pressing of industry requests for Big Data tools,
several ways of their fitting to this context were proposed in [50,134]; thus,
for Pattern Structures in classification setting, combination of lazy evaluation
with projection approximations of initial data, randomisation and parallelisa-
tion, results in reduction of algorithmic complexity to low degree polynomial.
This observations make it possible to apply pattern structures in text mining
and learning from large text collections [135]. Implementations of basic Pattern
Structures algorithms are available in FCART. 

Exercise 20. 1. Compose a small program, e.g. in Python, that enumerates all
pattern concepts from the movie recommender example directly by definition or
adapt CbO this end. 2. In case there is no possibility to perform 1., consider
the subtable of the first four users and the first four movies from the movie
recommender example. Find all pattern concepts by the definition. Build the
line diagram of the pattern concept lattice. 

However, Pattern Structures is not the only attempt to fit FCA to data
with more complex description than Boolean one. Thus, during the past years,
the research on extending FCA theory to cope with imprecise and incomplete
information made significant progress. The underlying model is a so called fuzzy
concepts lattice; there are several definitions of such a lattice, but the basic
assumption usually is that an object may posses attributes to some degree [136].
For example, in sociological studies age representation requires a special care:
a person being a teenager cannot be treated as a truly adult one on the first
day when his/her age exceeds a threshold of 18 years old (moreover, for formal
reasons this age may differ in different countries). However, it is usually the case
when we deal with nominal scaling; even ordinal scaling may lead to information
loss because of the chosen granularity level. So, we need a flexible measure of
being an adult and a teenage person at the same and it might be a degree lying in
[0,1] interval for each such attribute. Another way to characterise this imprecision
or roughness can be done in rough sets terms [137]. An interested reader is
invited to follow a survey on Fuzzy and Rough FCA in [138]. The correspondence
between Pattern Structures and Fuzzy FCA can be found in [139].

5.4 FCA-based Boolean Matrix Factorisation

Matrix Factorisation (MF) techniques are in the typical inventory of Machine

Learning ([125], chapter Features), Data Mining ([63], chapter Dimensionality
Reduction) and Information Retrieval ([1], chapter Matrix decompositions and
latent semantic indexing). Thus MF used for dimensionality reduction and fea-
ture extraction, and, for example, in Collaborative filtering recommender MF
techniques are now considered industry standard [140].
 Introduction to Formal Concept Analysis and Its Applications in IR 45

Among the most popular types of MF we should definitely mention Singular

Value Decomposition (SVD) [141] and its various modifications like Probabilis-
tic Latent Semantic Analysis (PLSA) [142] and SVD++ [143]. However, several
existing factorisation techniques, for example, non-negative matrix factorisation
(NMF) [144] and Boolean matrix factorisation (BMF) [51], seem to be less stud-
ied in the context of modern Data Analysis and Information Retrieval.
Boolean matrix factorisation (BMF) is a decomposition of the original matrix
I ∈ {0, 1}n×m , where Iij ∈ {0, 1}, into a Boolean matrix product P ◦ Q of binary
matrices P ∈ {0, 1}n×k and Q ∈ {0, 1}k×m for the smallest possible number of
k. Let us define Boolean matrix product as follows:
k
_
(P ◦ Q)ij = Pil · Qlj , (3)
l=1
W
where denotes disjunction, and · conjunction.
Matrix I can be considered a matrix of binary relations between set X of
objects (users), and a set Y of attributes (items that users have evaluated). We
assume that xIy iff the user x evaluated object y. The triple (X, Y, I) clearly
forms a formal context.
Consider a set F ⊆ B(X, Y, I), a subset of all formal concepts of context
(X, Y, I), and introduce matrices PF and QF :
 
1, i ∈ Al , 1, j ∈ Bl ,
(PF )il = (QF )lj = ,
0, i ∈
/ Al , 0, j ∈
/ Bl .

where (Al , Bl ) is a formal concept from F .

We can consider decomposition of the matrix I into binary matrix product
PF and QF as described above. The following theorems are proved in [51]:

Theorem 2. (Universality of formal concepts as factors). For every I there is

F ⊆ B(X, Y, I), such that I = PF ◦ QF .

Theorem 3. (Optimality of formal concepts as factors). Let I = P ◦Q for n×k

and k × m binary matrices P and Q. Then there exists a set F ⊆ B(X, Y, I) of
formal concepts of I such that |F| ≤ k and for the n × |F| and |F| × m binary
matrices PF and QF we have I = PF ◦ QF .

Example 17. Transform the matrix of ratings described above by thresholding

(geq3), to a Boolean matrix, as follows:
 
1110000
1 1 1 1 1 0 0
 
0 0 0 1 1 0 0
 0 0 0 1 1 1 1  = I.
 
 
0 0 0 0 0 1 1
0000011
46 Dmitry I. Ignatov

The decomposition of the matrix I into the Boolean product of I = AF ◦ BF is

the following:
   
1110000 100
1 1 1 1 1 0 0 1 1 0  
  
0 0 0 1 1 0 0 0 1 0
 1110000
0 0 0 1 1 1 1 = 0 1 1 ◦ 0 0 0 1 1 0 0 .
     
  
0 0 0 0 0 1 1 0 0 1
 0000011
0000011 001

Even this tiny example shows that the algorithm has identified three factors
that significantly reduces the dimensionality of the data. 

There are several algorithms for finding PF and QF by calculating formal

concepts based on these theorems [51]. Thus, the approximate algorithm (Al-
gorithm 2 from [51]) avoids computation of all possible formal concepts and
therefore works much faster than direct approach by all concepts generation. Its
running time complexity in the worst case yields O(k|G||M |3 ), where k is the
number of found factors, |G| is the number of objects, |M | is the number of
attributes.
As for applications, in [120,121], FCA-based BMF was used as a feature
extraction technique for improving the results of classification. Another example
closely relates to IR; thus, in [145,94] BMF demonstrated comparable results to
SVD-based collaborative filtering in terms of MAE and precision-recall metrics.
Further extensions of BMF to triadic and n-ary data were proposed in [146]
and [147], respectively (the last one in not FCA-based)

5.5 Case study: admission process to HSE university

in this case study we reproduce results of our paper from [52]. Assuming probable
confusion of the Russian educational system, we must say a few words about the
National Research University Higher School of Economics19 and its admission
process.
Nowadays HSE is acknowledged as a leading university in the field of eco-
nomics, management, sociology, business informatics, public policy and political
sciences among Russian universities. Recently a number of bachelor programmes
offered by HSE has been increased. In the year 2010 HSE offered 20 bachelor
programmes. We consider only bachelor programmes in our investigation.
In order to graduate from school and enter a university or a college every
Russian student must pass a Unified State Exam (Russian transcription: EGE),
similar to US SAT–ACT or UK A-Level tests. During 2010 admission to U-HSE,
entrants were able to send their applications to up to three programmes simul-
taneously. Some school leavers (major entrants of HSE bachelor programmes)
chose only one programme, some chose two or three. Then entrants had to choose
only one programme to study among successful applications.
19
http://www.hse.ru/en/
 Introduction to Formal Concept Analysis and Its Applications in IR 47

We used data representing admission to HSE in 2010. It consists of informa-

tion about 7516 entrants. We used mainly information about programmes (up
to three) to which entrants apply20 . Exactly 3308 entrants successfully applied
at least to one programme, but just 1504 become students. Along with this data
we also used the data of entrants’ survey (76% of entire assembly).
Further in the paper we mostly used data for the Applied Mathematics and
Informatics programme to demonstrate some results. The total number of ap-
plications to the Applied Mathematics and Informatics programme was 843, of
which 398 were successful but only 72 of them were actually accepted into the
program. It might seem confusing only 72 out of 398 eligible prospective students
decided to enroll, but since the admission process was set up in two stages, and
at each stage only 72 entrants were eligible to attend the program, some of them
decided to go for a different programme or university. As a result, the number of
entrants whose applications were successful in any way came down to 398. Such
situation is typical for all the bachelor programmes at HSE.
FCA requires object-attribute data. In our case objects are entrants and pro-
grammes they apply to are attributes. Together they are treated as a context.
A series of contexts were constructed. Namely, we built a context for every pro-
gramme where objects were entrants applying to that programme and attributes
were other programmes they applied to. We built a separate context for every
programme because it is meaningless to consider all programmes at once as pro-
grammes are very different in size and the resulting lattice would represent only
the largest of them.
Likewise, we built a context for every programme where objects were entrants
and attributes were programmes to which entrants successfully applied as well
as the programmes that the entrants decided to enroll into, including those at
other universities.
These contexts were then used to build concept lattices. Since the resulting
lattices had too complicated a structure to interpret, we filtered concepts by their
extent size (extent size is the number of objects, in our case it is the number
of entrants), thus remaining concepts express only some of the more common
patterns in entrants decisions.
To which programmes entrants often apply simultaneously? Trying to answer
this question for every programme, we built diagrams21 similar to figure 16.
Such diagrams help us to reveal common patterns in entrants choices. Typical
applications of FCA imply building formal concept lattices discussed earlier, but
here we filter concepts by extent size to avoid complexity caused by noise in the
data. Thus the order on remaining concepts is no longer a lattice, it is a partial
order. Meaning of the labels on the diagram is obvious. A label above a node is a
programme, a label below a node is a percent of entrants to Applied Mathematics
20
HSE is a state university, thus most of student places are financed by government.
In this paper we consider only such places.
21
As any other data mining technique FCA implies an intensive use of software. All
diagrams mentioned in this paper have been produced with meud (https://github.
com/jupp/meud-wx).
48 Dmitry I. Ignatov

Applied Mathematics and Informatics

Mathematics Sociology
100
Statistics
10 Software Engineering 1
Public Administration
Economics
1
World Economy
70 Business Informatics Management Logistics
1
4
5 2
30 2 4
1
19
1
1 2

Fig. 16. Other programmes which entrants of Applied Mathematics and Informatics
programme also applied.

and Informatics programme who also applied to programmes connected to a

node from above. For example, the most left and bottom node on the diagram
means that five percent of applied math’s entrants also apply to Mathematics
and Software Engineering. Then if we look at the nodes above the current node
we may notice that ten percent Applied Mathematics and Informatics applicants
also apply to Mathematics programme, and 70 percent also applied to Software
Engineering.
Now let us try to interpret some knowledge unfolded by the diagram in
figure 16. 70 percent of entrants who applied to Applied Mathematics and In-
formatics also apply to Software Engineering. The same diagram for Software
Engineering states that 80 percent of Software Engineering applicants also ap-
ply to Applied Mathematics and Informatics. How this fact can be explained?
Firstly it can easily be explained by the fact that these two programmes require
to pass the same exams. Therefore there were not any additional obstacles to
apply to both programmes simultaneously. Another possible explanation is that
it is uneasy for entrants to distinguish these two programmes and successful
application to any of them would be satisfactory result.
Analysing diagrams of other programmes’ applications we found that equiva-
lence of required exams is probably the most significant reason to apply to more
than one programme.

Entrants’ ”Efficient” choice. If an entrant successfully applied to more than

one bachelor programme he or she must select a programme to study. Unlike
the previous case, entrants have to select exactly one programme which gives us
more precise information about entrants preferences. For that reason we define
this situation as an efficient choice, efficient in the sense of more expressive about
true entrants preferences.
Figure 17 presents the efficient choice of entrants to Applied Mathematics
and Informatics programme. The meaning of diagram labels is almost the same
as in Fig. 16. Programmes without plus sign (+) are successful applications, pro-
grammes with preceding plus sign are programmes chosen to study by entrants.
 Introduction to Formal Concept Analysis and Its Applications in IR 49

Applied Mathematics and Informatics

Logistics 100
Management
Business Informatics
- Other - Software Engineering
Economics
6
Mathematics 3
32
69 46
3
+ Applied Mathematics and Informatics
8
+ Business Informatics + Software Engineering
3 4
18
5 22 5
17

3
33
5
2 3
7
2
12

Fig. 17. “Efficient” choice of entrants to Applied Mathematics and Informatics pro-
gramme.

Label “- Other -” means that the entrant canceled his application preferring
another university or not to study this year altogether.
Together with diagram in Fig. 16 1 this diagram provides us with more pre-
cise knowledge about preferences of entrants to the Applied Mathematics and
Informatics programme. More than two thirds of entrants who successfully apply
to the Applied Math programme nevertheless prefer to study at another univer-
sity. Whereas just 18 percent of successful applicants then become students on
the Applied Mathematics and Informatics programme. Exactly 5 percent prefer
to study Software Engineering and 5 percent of entrants who choose Applied
Mathematics and Informatics also successfully applied to Software Engineer-
ing. It can be interpreted as equality of entrants preferences concerning these
two programmes. Additionally, 5 percent prefer Business Informatics and only
two percent of entrants who prefer Applied Mathematics and Informatics also
successfully apply to Business Informatics, therefore in the pair Business Infor-
matics and Applied Mathematics and Informatics the latter one is less preferable
by entrants.
Here we should note that the sum of nodes percents with labels containing
plus sign and node “- Other -” must equal to 100%, however here it does not
because we excluded some nodes during filtering.
We built diagrams of “efficient” choice for every programme. Analysis of these
diagrams helps us to recognise some relations between programmes in terms of
entrants preferences. For example, some programmes in most cases is rather
backup than actual entrants preference. Some programmes are close to each
other by subject of study, these relations are also expressed by diagrams. With
help of formalised survey data we found some possible factors of entrants’ choice
among some particular programmes. These knowledge can help our university to
50 Dmitry I. Ignatov

understand entrants’ attitude to its undergraduate programmes and thus correct

the structure and positioning of them.
Another Educational data mining case includes analysis of student achieve-
ments in two subsequent year for the same group by means of grading data [148].

5.6 Machine learning exercises with JSM-method in QuDA

QuDA was developed in early 2000s as “a software environment for those who
want to learn Data Mining by doing” at the Intellectics group of the Darm-
stadt Technical University of Technology [149,150,53]. It includes various tech-
niques, such as association rule mining, decision trees and rule-based learning,
JSM-reasoning (including various reasoning scheme [151]), Bayesian learning,
and interesting subgroup discovery. It also provides the experimenter with error
estimation and model selection tools as well several preprocessing and post-
processing utilities, including data cleansing tools, line diagrams, visualisation
of attribute distributions, and a convenient rule navigator, etc. It was mostly
aimed to support scientific and teaching activities in the field of Machine Learn-
ing and Data Mining. However, since QuDA has open architecture and support
the most common data formats as well as the Predictive Model Markup Lan-
guage (PMML)22 , it can be easily integrated into a working Data Mining circle.
Originally, it was an acronym for “Qualitative Data Analysis”. Now, since QuDA
finally includes many quantitative methods integrated into it from WEKA23 , this
name is a backronym 24 since it has lost its original meaning.

Exercise 21. Download QuDa 25 . Refer to QuDa’s manual [149] for details and
prepare the credit scoring context in csv format for opening in the QuDA en-
vironment. Perform nominal scaling of attributes and apply JSM classifier with
basic setup. Compare the obtained rules with the hypotheses obtained manually.


Exercise 22. For zoo dataset available with QuDa (or some other dataset that
suitable for classification from UCI ML repository26 ), perform nominal scaling
and comparison of JSM-classification against all available methods 1) by splitting
data into 80:20 training-to-test sample size ration 2) by 10-fold cross-validation.
Compare learning curves and confusion matrices. Identify all non-covered ex-
amples by JSM-method. Change scaling type for attribute “number of legs”.
Reiterate comparison and check which methods have improved their classifica-
tion quality. 
22
http://www.dmg.org/
23
http://www.cs.waikato.ac.nz/ml/weka/
24
http://en.wikipedia.org/wiki/Backronym
25
http://sourceforge.net/projects/quda/; its alternative compilation for RuSSIR
2014 is vaialable at http://bit.ly/QuDA4RuSSIR2014
26
http://archive.ics.uci.edu/ml/datasets.html
 Introduction to Formal Concept Analysis and Its Applications in IR 51

6 FCA in Information Retrieval and Text Mining

Lattice-based models and FCA itself are not mainstream directions of modern
IR; they attracted numerous researchers because of their interpretability and
human-centerdness, but their intrinsic complexity is a serious challenge to make
them working on a Web scale.
Thus, from early works on Information Retrieval it is known that usage of a
lattice as a search space requires treatment of the enormous number of subsets
of documents: 10310,100 for a collection of one million documents [152]. At that
time it was rather natural in library classification domain to consider documents
and their categories, which may form requests as a combination of simple logical
operations like AND, NOT, and OR [153]. Thus, Mooers considered transfor-
mations T : P → L, where P is the space of all possible document descriptors
and L is the space of all possible document subsets [152]. Thus, T retrieves the
largest set of documents from L according to a query (prescription) from P .
At that period in Soviet Russia, All-Soviet Institute for Scientific and Tech-
nical Information (VINITI) was organised to facilitate information interchange
and fulfill growing scientific needs in cataloging and processing of scientific pub-
lications. Around the mid 1960s, Yulii A. Shreider, one of the leading researchers
of VINITI, considered the problem of automatic classification of documents and
their retrieval by means of a model featuring a triple (M, L, f ), where M is a
set of documents, L is a set of attributes and f : M → 2L maps each document
to a set attributes from L [154]. There, similarity of two documents was defined
via non-emptiness of the intersection of their descriptions f (d1 ) ∩ f (d2 ). In that
paper, Shreider mentioned the relevance of lattices to problems of document
classification and retrieval, where he also cited the work of Soergel [155] on this
issue.
Thus, these two introduced mappings, T and f highly resemble to conven-
tional prime operators in FCA for the context of documents and their attributes
(keywords, terms, descriptors) with “document-term containment” relation. In
the middle of 80s, Godin et al. [156] proposed a lattice-based retrieval model
for database browsing, where objects (documents, e.g. course syllabi) were de-
scribed by associated keywords. The resulting (in fact, concept) lattice used for
navigation by query modification using its generality/specificity relation.
In 90-s several FCA-based IR models and systems appeared, the reviews
can be found [157,158]. Thus, in [157], Carpineto and Romano classified main
IR problems that can be solved by FCA means through review of their own
studies by the year 2005. Uta Priss described a current state of FCA for IR
domain [158] by the year 2004. Recently, a survey on FCA-based systems and
methods for IR including prospective affordances was presented at FCA for
IR workshop at ECIR 2013 [159], and our colleagues, Codocedo and Napoli,
taking an inspiration, are summarising the latest work on the topic in a separate
forthcoming survey.
Below, we shortly overview our own study on applying FCA-based IR meth-
ods for describing the state-of-the-art in FCA for IR field. The rest topics are
52 Dmitry I. Ignatov

spread among the most representative examples of FCA-based IR tasks and

systems including the summary of the author’s experience.

– Text Mining scientific papers: a survey on FCA-based IR applications [37]

– FCA-based meta-search engines (FOOCa, SearchSleuth, Credo etc.) [83,15]
– FCA-based IR visualisation [15] and navigation (ImageSleuth, Camelis [160])
– FCA in criminology: text mining of police reports [37]
– FCA-based approach for advertising keywords in web search [96]
– FCA-based Recommender Systems [145]
– Triadic FCA for IR-tasks in Folksonomies [161]
– FCA-based approach for near-duplicate documents detection [162,81]
– Exploring taxonomies of web site users [163]
– Concept-based models in Crowdsourced platforms: a recommender system
of like-minded persons, antagonists and ideas [164]

6.1 Text Mining scientific papers: a survey on FCA-based IR

applications

In [4], we visually represented the literature on FCA and IR using concept lat-
tices, in which the objects are the scientific papers and the attributes are the
relevant terms available in the title, keywords and abstract of the papers. We
developed an IR tool with a central FCA component that we use to index the
papers with a thesaurus containing terms related to FCA research and to gen-
erate the lattices. It helped us to zoom in and give an extensive overview of 103
papers published between 2003 and 2009 on using FCA in information retrieval.
Information Retrieval
web services
browsing

software
mining

FCA

P aper1 × × × ×
P aper2 × ××
P aper3 × × ×
P aper4 × × ×
P aper5 × ××

We developed a knowledge browsing environment CORDIET to support our

literature analysis process. One of the central components of our text analysis
environment is the thesaurus containing the collection of terms describing the
different research topics. The initial thesaurus was constructed based on expert
prior knowledge and was incrementally improved by analyzing the concept gaps
 Introduction to Formal Concept Analysis and Its Applications in IR 53

and anomalies in the resulting lattices. The layered thesaurus contains multiple
abstraction levels. The first and finest level of granularity contains the search
terms of which most are grouped together based on their semantic meaning to
form the term clusters at the second level of granularity. The papers downloaded
from the Web were converted to plain text and the abstract, title and keywords
were extracted. The open source tool Lucene27 was used to index the extracted
parts of the papers using the thesaurus. The result was a cross table describing
the relationships between the papers and the term clusters or research topics
from the thesaurus. This cross table was used as a basis to generate the lattices.
The most relevant scientific sources sources that were used in the search
for primary studies contain the work published in those journals, conferences
and workshops which are of recognized quality within the research com-munity.
These sources are: IEEE Computer Society, ACM Digital Library, Sciencedi-
rect, Springerlink, EBSCOhost, Google Scholar, Conference repositories: ICFCA,
ICCS and CLA conferences. Other important sources such as DBLP or CiteSeer
were not explicitly included since they were indexed by some of the mentioned
sources (e.g. Google Scholar). In the selected sources we used various search
terms including “Formal Concept Analysis”, “FCA”, “concept lattices”, “Infor-
mation Retrieval”. To identify the major categories for the literature survey we
also took into account the number of citations of the FCA papers at CiteseerX.
The efficient retrieval of relevant information is promoted by the FCA repre-
sentation that makes the inherent logical structure of the information transpar-
ent. FCA can be used for multiple purposes in IR [15,158]. First, FCA provides
an elegant language for IR modeling and is an interesting instrument for brows-
ing and automatic retrieval through document collections. Second, FCA can
also support query refinement, ranking and enrichment by external resources.
Because a document-term lattice structures the available information as clus-
ters of related documents which are partially ordered, lattices can be used to
make suggestions for query enlargement in cases where too few documents are
retrieved and for query refinement in cases where too many documents are re-
trieved. Third, lattices can be used for querying and navigation supporting rele-
vance feedback. An initial query corresponds to a start node in a document-term
lattice. Users can then navigate to related nodes. Further, queries are used to
“prune” a document-term lattice to help users focus their search (Carpineto et
al. 1996b). For many purposes, some extra facilities are needed such as processing
large document collections quickly, allowing more flexible matching operations,
allowing ranked retrieval and give contextual answers to user queries. The past
years many FCA researchers have also devoted attention to these issues.
86 % of the papers on FCA and information retrieval are covered by the
research topics in Fig. 6.1. Further, in our study, we intuitively introduced the
process of transforming data repositories into browsable FCA representations
and performing query expansion and refinement operations. Then we considered
28 % of papers on using FCA for representation of and navigation in image,
service, web, etc. document collections. Defining and processing complex queries
27
https://lucene.apache.org/core/
54 Dmitry I. Ignatov

Fig. 18. Lattice containing 103 papers on using FCA in IR

covered 6% of the papers and was described as well. The review of papers on
contextual answers (6% of papers) and ranking of query results (6% of papers)
concluded the case-study.

Knowledge representation and browsing with FCA In 28% of the 103 selected
papers, FCA is used for browsing and navigation through document collections.
In more than half of these papers (18% of total number of papers), a combination
of navigation and querying based on the FCA lattices is pro-posed. Annotation
of documents and finding optimal document descriptors play an important role
in effective information retrieval (9% of papers). All FCA-based approaches for
information retrieval and browsing through large data repositories are based on
the same underlying model. We first have the set D containing objects such as
web pages, web services, images or other digitally available items. The set A
of attributes can consist of terms, tags, descriptions, etc. These attributes can
be related to certain objects through a relation I ⊆ D × A which indicates the
terms, tags, etc. can be used to describe the data elements in D. This triple
(D, A, I) is a formal context from which the concept lattice can be created.

Query result improvement with FCA Search engines are increasingly being used
by amongst others web users who have an information need. The intent of a
concept corresponds to a query and the extent contains the search results. A
query q features a set of terms T and the system returns the answer by evaluating
 Introduction to Formal Concept Analysis and Its Applications in IR 55

T 0 . Upon evaluating a query q the system places itself on the concept (T 0 , T 00 )

which becomes the current concept c. For example, in Fig. 19, the intent of the
current concept Bc = {t1 , t2 , t3 , t4 , t5 } and the extent of the current concept
Ac = {d8 , d9 }, where t stands for term and d stands for word. Since a query
provided by a user only approximates a user’s need, many techniques have been
developed to expand and refine query terms and search results. Query tuning is
the process of searching for the query that best approximates the information
need of the user. Query refinements can help the user express his original need
more clearly. Query refinement can be done by going to a lower neighbor of
the current concept in the lattice by adding a new term to the query items. In
minimal conjunctive query refinement the user can navigate for example
to a subconcept ((Bc ∪ {t})0 , (Bc ∪ {t})00 ) by adding term t.

((Ac ∪ {d})00 , (Ac ∪ {d})0 )

({d8 , d9 }, {t1 , t2 , t3 , t4 , t5 })

((Bc ∪ {t})0 , (Bc ∪ {t})00 )

Fig. 19. Query modification in a concept lattice: a fish-eye view

Query enlargement, i.e. retrieving additional relevant web pages, can be per-
formed by navigating to an upper neighbor of the current concept in the lattice
by removing a term from the query items. The user can navigate for example
to a superconcept ((Bc ∪ {d})00 , (Bc ∪ {d})0 ) by adding document d. The com-
bination of subsequent refine and expand operations can be seen as navigation
through the query space. Typically, navigation and querying are two completely
separate processes, and the combination of both results in a more flexible and
user-friendly method. These topics are investigated in 8 % of the IR papers. See
the comprehensive survey on query refinement in [165].

Concept lattice based ranking. 6% of IR papers in our study are devoted concept
lattice ranking.
Below we explain the concept lattice based ranking (CLR) proposed in [166]
and compared with hierarchical clustering based (HCR) and best-first matching
(BFR) rankings. The experiments with two public benchmark collections showed
that CLR outperformed the two competitive methods when the ranked docu-
ments did no match the query and was comparable to BMF and better than
HCR in the rest cases.
Let (D, T, I) be a document context, where D is the set of documents,
T is the set of their terms, and I ⊆ D × T . Consider the ordered set of
56 Dmitry I. Ignatov

all concepts (L(D, T, I), ≺) with the nearest neighbour relation ≺, i.e., for
c1 , c2 ∈ L(D, T, I), c1 ≺ c2 iff c1  c2 or c1 ≺ c2 . Then distance between
concepts c1 and c2 is defined as the least natural number n as follows:
∃ci1 , . . . , cin ∈ L(D, T, I) such that c1 = ci0 ≺ ci1 . . . ≺ cin = c2 .
For each query q and d1 , d2 ∈ D, d1 is ranked higher than d2 if the distance
between (d001 , d01 ) and (q 00 , q 0 ) is shorter than than the distance between (d002 , d02 )
and (q 00 , q 0 ). It means that we need less query transformations to obtain q from
d1 than to do that from d2 .
As the author of CLR admitted in [15], CLR is sensitive to addition new
documents into the search collection. Further analysis of usage of concept neig-
bourhood based similarity for IR needs is given in [84].

Example 18. In Fig. 20 we provide an example of concept lattice based ranking

for the previous context of papers and their terms. The underlying query is the
conjunction of two terms: “browsing, FCA”. The query concept is as follows:

({p1 , p4 }, {browsing, F CA, sof tware}).

The resulting ranking yields p4 < p1 < p2 < p3 = p5 . A curious reader may
admit that concepts with the same ranks lie in concentric circles around the
query concept at the same distance. Obviously, for concepts from the same such
circle we need their subsequent ranking, e.g. by best-match ranking via dot
product of the document and query profiles based on term frequency.

FCA
2
software
1
web services
3 IR mining
browsing
3 2 0 p4

p5 3 p3 3 2 p2 1 p1

Fig. 20. Concept lattice based ranking for the query “browsing, FCA”; the distance
values are given inside the corresponding circles.

An interested reader can find the rest sections in our survey:

 Introduction to Formal Concept Analysis and Its Applications in IR 57

– Web and email retrieval (partially covered in Section 6.2);

– Image, software and knowledge base retrieval (partially covered in Section 6.3);
– Defining and processing complex queries with FCA;
– Domain knowledge in search results: contextual answers & ranking.

6.2 FCA-based IR visualisation and meta-search engines

From the beginning 2000s many independent IR developers proposed so called

meta-search engines also known as search results clustering engines. To name
a few, two project are still alive Carrots228 and Nigma.ru29 . See summarising
survey on Web clustered search by Carpineto et al. in [167].
FCA has been used as the basis for many web-based knowledge browsing sys-
tems developed during the past years. Especially its comprehensible visualisation
capabilities seem to be of interest to RuSSIR audience. The results returned by
web search engines for a given query are typically formatted as a list of URLs
accompanied by a document title and a a snippet, i.e. a short summary of the
document. Several FCA-based systems were developed for analyzing and explor-
ing these search results. CREDO [168], FooCA [83] and SearchSleuth [169,170]
build a context for each individual query which contains the result of the query
as objects and the terms found in the title and summary of each result as at-
tributes.
The CREDO system30 then builds an iceberg lattice which is represented as
a tree and can be interactively explored by the user.
FooCA31 shows the entire formal context to the user and offers a great degree
of flexibility in exploring this table using the ranking of attributes, selecting the
number of objects and attributes, applying stemming and stop word removal
etc.
SearchSleuth does not display the entire lattice but focuses on the search
concept, i.e. the concept derived from the query terms. The user can easily
navigate to its upper and lower neighbors and siblings. Nauer et al.[171] also
propose to use FCA for iteratively and interactively analyzing web search re-
sults. The user can indicate which concepts are relevant and which ones are
not for the retrieval task. Based on this information the concept lattice is dy-
namically modified. Their research resulted in the CreChainDo system32 . Kim
et al. [172] presented the FCA-based document navigation system KAnavigator
for small web communities in specialized domains. Relevant documents can be
annotated with keywords by the users. Kim et al. [173] extended the search func-
tionality by combining lattice-based browsing with conceptual scales to reduce
the complexity of the visualisation. Cigarran et al. [174] present the JBrainDead
IR System which combines free-text search with FCA to organise the results
28
http://search.carrot2.org/
29
http://www.nigma.ru/
30
http://credo.fub.it/
31
http://www.bjoern-koester.de/
32
http://orpailleur.loria.fr/index.php/CreChainDo
58 Dmitry I. Ignatov

Fig. 21. An example of CREDO’s web search interface

Fig. 22. An example of FooCA’s web search interface. It processes the results of search
queries to Yahoo or Google and organises them into the interactive cross-table.
 Introduction to Formal Concept Analysis and Its Applications in IR 59

Fig. 23. An example of SearchSleuth’s web interface. It processes the results of search
queries to Yahoo. Passing to more general (more specific) categories is done by clicking
-term(+term).
60 Dmitry I. Ignatov

of a query. Cole et al. [175] discuss a document discovery tool named Concep-
tual Email Manager (CEM) which is based on FCA. The program allows users to
navigate through emails using a visual lattice. The paper also discusses how con-
ceptual ontologies can support traditional document retrieval systems and aid
knowledge discovery in document collections. The development of this software
is based on earlier research on retrieval of information from semi-structured texts
([176,177]). Building further on this work is the Mail-Sleuth software (Eklund
et al. [178]) which can be used to mine large email archives. Eklund et al. [179]
use FCA for displaying, searching and navigating through help content in a help
system. Stojanovic [180] presents an FCA-based method for query refinement
that provides a user with the queries that are “nearby” the given query. Their
approach for query space navigation was validated in the context of searching
medical abstracts. Stojanovic [180] presents the SMART system for navigation
through an online product catalog. The products in the database are described
by elements of an ontology and visualized with a lattice, in which users can nav-
igate from a very general product-attribute cluster containing a lot of products
to very specific clusters that seem to contain a few, but for the user highly rele-
vant products. Spyratos et al. [181] describe an approach for query tuning that
integrates navigation and querying into a single process. The FCA lattice serves
for navigation and the attributes for query formulation. Le Grand et al. [182]
present an IR method based on FCA in conjunction with semantics to provide
contextual answers to web queries. An overall lattice is built from tourism web
pages. Then, users formulate their query and the best-matching concepts are
returned, users may then navigate within the lattice by generalizing or refining
their query. Eklund et al. [183] present AnnotationSleuth to extend a standard
search and browsing interface to feature a conceptual neighborhood centered on
a formal concept derived from curatorial tags in a museum management system.
Cigarran et al. [184] focus on the automatic selection of noun phrases as docu-
ments descriptors to build an FCA based IR system. Automatic attribute selec-
tion is important when using FCA in a free text document retrieval framework.
Optimal attributes as document descriptors should produce smaller, clearer and
more browsable concept lattices with better clustering features. Recio-Garcia et
al. [185] use FCA to perform semantic annotation of web pages with domain
ontologies. Similarity matching techniques from Case Based Reasoning can be
applied to retrieve these annotated pages as cases. Liu et al. [186] use FCA to
optimise a personal news search engine to help users obtain the news content
they need rapidly. The proposed technique combines the construction of user
background using FCA, the optimisation of query keywords based on the user’s
background and a new layout strategy of search results based on a “Concept
Tree”. Lungley et al. [187] use implicit user feedback for adapting the underly-
ing domain model of an intranet search system. FCA is used as an interactive
interface to identify query refinement terms which help achieve better document
descriptions and more browsable lattices.
 Introduction to Formal Concept Analysis and Its Applications in IR 61

6.3 FCA-based image retrieval and navigation

FCA-based IR visualisation [15] and navigation (ImageSleuth, Camelis [160])

Ahmad et al. [188] build concept lattices from descriptions associated to
images for searching and retrieving relevant images from a database. In the
ImageSleuth project [189], FCA was also used for clustering of and navigation
through annotated collections of images. The lattice diagram is not directly
shown to the user. Only the extent of the present concept containing thumb-
nails, the intent containing image descriptions and a list of upper and lower
neighbors is shown. In Ducrou [190], the author built an information space from
the Amazon.com online store and used FCA to discover conceptually similar
DVDs and explore their conceptual neighborhood. The system was called DVD-
Sleuth. Amato et al. [191] start from an initial image given by the user and use
a concept lattice for retrieving similar images. The attributes in this lattice are
facets, i.e. an image similarity criterion based on e.g. texture, color or shape.
The values in the context indicate for each facet how similar an image in the
database is with respect to the user provided initial image. By querying, the
user can jump to any cluster of the lattice by specifying the criteria that the
sought cluster must satisfy. By navigation from any cluster, the user can move
to a neighbor cluster, thus exploiting the ordering amongst clusters.
In [160] Ferre et al. proposed to use so called Logical Information Systems
(LIS) for navigation through photo collections33 . In fact LIS, similarly to Pattern
Structures, exploit partially ordered object descriptions but expressed as logical
formulas.

– location: Nizhniy Novgorod v Russia

– date: date = 18 Aug 2014 v date in Aug 2014 .. Jul 2015
– event: event is “summer school RuSSIR” v event contains “summer school”

Further it was extended for work with document collections[192]. Since Camelis
uses lattice-based navigation and queriying by formulas, it overcomes current
drawbacks of tree-like navigation imposed by current restrictions of file-systems.
Recently, the previous studies of Eklund et al. [84] in organising navigation
through annotated collections of images in virtual museums resulted in an iPad
application that allows users to explore an art collection via semantically linked
pathways that are generated using Formal Concept Analysis34 . In fact navigation
in this application is organised by showing context and relationships among
objects in a museum collection.

6.4 FCA in criminology: text mining of police reports

In [37], we proposed an iterative and human-centred knowledge discovery method-

ology based on FCA. The proposed approach recognises the important role of the
33
Camelis, http://www.irisa.fr/LIS/ferre/camelis/
34
“A place for art”, https://itunes.apple.com/au/app/a-place-for-art/
id638054832?mt=8
62 Dmitry I. Ignatov

domain expert in mining real-world enterprise applications and makes use of spe-
cific domain knowledge, including human intelligence and domain-specific con-
straints. Our approach was empirically validated at the Amsterdam-Amstelland
police to identify suspects and victims of human trafficking in 266,157 suspicious
activity reports. Based on guidelines of the Attorney Generals of the Netherlands,
we first defined multiple early warning indicators that were used to index the
police reports.

Example 19. This is an example of a police report where some indicator-words

are highlighted that are used for its contextual representation.
Report 1:
On the night of 23 of March

Former Eastern Europe

2008 we stopped a car with a

Information Retrieval
Bulgarian license plate for
routine motor vehicle inspection.

Expensive cars
It was a Mercedes GLK with
license plate BL XXX. The car

Prostitutes
Id-papers
was driving around in circles

Vacation
in a prostitution area. On the
backseat of the car we noticed
two well dressed young girls. We
asked for their identification Report1 × × × × × ×
papers but they did not speak Report2 × × × × ×
English nor Dutch. The driver Report3 × × × ×
of the car was in possession of Report4 × ×
their papers and told us that Report5 × ×
they were on vacation in the
Netherlands for two weeks etc.


Our method based on FCA consists of four main types of analysis which are
carried out as follows:

1. Concept exploration of the forced prostitution problem of Amsterdam: In

Poelmans et al. [193], this FCA-based approach for automatically detecting
domestic violence in unstructured text police reports is described in detail.
2. Identifying potential suspects: concept lattices allow for the detection of
potentially interesting links between independent observations made by dif-
ferent police officers.
3. Visual suspect profiling: some FCA-based methods such as temporal concept
analysis (TCA) were developed to visually represent and analyse data with a
temporal dimension [194]. Temporal concept lattices were used in Elzinga et
al. [195] to create visual profiles of potentially interesting terrorism subjects.
Elzinga et al. [196] used TCA in combination with nested line diagrams to
analyse pedophile chat conversations.
4. Social structure exploration: concept lattices may help expose interesting
persons related to each other, criminal networks, the role of certain sus-
 Introduction to Formal Concept Analysis and Its Applications in IR 63

pects in these networks, etc. With police officers we discussed and compared
various FCA-based visualisation methods of criminal networks.

In our investigations we also used the model that was developed by Bullens
and Van Horn [197] for the identification of loverboys who typically force girls of
Dutch nationality into prostitution. Loverboys use their love affair with a woman
to force her to work in prostitution. Forcing girls and women into prostitution
through a loverboy approach is seen as a special kind of human trafficking in the
Netherlands (article 250a of the code of criminal law). This model is a resource
used by the Amsterdam-Amstelland police during the trainings of police officers
about this topic. A typical loverboy approach consists of three main phases which
give rise to corresponding indicators:
1. Preparatory activities to recruit girls.
2. Forcing her into prostitution.
3. Keeping the girl in prostitution by emotional dependence or social isolation.

The pimp will also try to protect his organisation.

In our data-set, there were three reports available about girl H. The reports
about this girl led us to the discovery of loverboy suspect B. The first report
(26 November 2008) contains the notification of the police by a youth aid organ-
isation in Alkmaar about girl H. They report a suspicious tattoo on her wrist
containing the name B. This B refers to her boyfriend who carries the same first
name, is 30 years old, and is of Surinamian origin. The second report was written
by a police officer who works in the red light district and knows many women
working in brothels or behind the windows. During a patrol he saw H working
as a prostitute, had a conversation with her, and observed suspicious that his
included in the report. The next report contains four suspicious facts recorded
by the officer. First, an unbelievable story why she works as a prostitute: a bet
between girlfriends if someone would dare to work as a prostitute. Second, the
tattoos of which one tattoo is mentioned in the first report (B) and a new one
on her belly. Third, the injuries, she has scratches on her arm (possibly from a
fight) and burns on her leg.
According to the victim, she has dropped a hot iron on her leg and had an
accident with a gourmet set. Fourth is the observation of making long working
days. The third document (21 December 2008) showed an observation of the
victim walking with the possible suspect. In this document the police officer re-
ports that he saw the victim and a man walking close to each other. The police
officer knows the man and knows that he is active in the world of prostitution.
When the man saw the officer, he immediately took some distance of the victim.
As soon as they have passed the officer, they walk close together and into a
well-known street where prostitutes work behind the windows. The first name
of the person is B, the same name which is tattooed on the victims wrist, and
the description of the person is about the same as described by the youth aid or-
ganisation. This information signals that the man is the possible loverboy of the
victim. The three reports together give serious presumptions of B being a lover-
boy with H being the victim. The next step is investigating B. We need serious
64 Dmitry I. Ignatov

indications that B is really involved in forced prostitution. Twelve observational

reports were found for B and the resulting lattice is shown in Figure 24.

Fig. 24. Line diagram for the report context of loverboy suspect B

Investigating these reports shows that he frequently visits the red light dis-
trict and has strong relationships with other pimps. One of these pimps is the
suspect of another loverboy case. From the six observations where B was seen in
the red light district, four are violence related, including the observation of Hs
suspicious burn wounds. The other violence-related observations are situations
of fights with customers who are unwilling to leave or pay. Such violence-related
observations are related to pimps who want to protect their prostitutes from
customers and competing gangs. Since in the Netherlands, prostitution is legal,
each prostitute has the right to ask the police to protect her. The violence ob-
servations of the suspect strengthened the suspicion of B being the pimp of H.
Moreover, we found another girl R who was also a potential victim of him. These
indications were enough to create a summary report and send a request for using
further investigation techniques to the public prosecutor.
Using concept lattices, we revealed numerous unknown human trafficking and
loverboy suspects. Indepth investigation by the police resulted in a confirmation
of their involvement in illegal activities resulting in actual arrestments been
made. This approach was embedded into operational policing practice and is
now successfully used on a daily basis to cope with the vastly growing amount
of unstructured information.
 Introduction to Formal Concept Analysis and Its Applications in IR 65

There are other FCA-based studies in criminology, for example, terrorist

activity modeling and analysis [198] and developments of lattice-based access
policies for information systems [199,200].

6.5 FCA-based approach for advertising keywords in web search

Online advertising by keywords matching is bread and butter of modern web

search companies like Google and Yandex. For our experimentation we used data
of US Overture [201] (now, a part of Yahoo), which were first transformed in the
standard context form. We consider the following context: KF T = (F, T, IF T ),
where F is the set of advertising firms (companies), T is the set of advertising
terms, or phrases, f IF T t means that firm f ∈ F bought advertising term t ∈ T .
In the context |F | = 2000, |T | = 3000, |IF T | = 92345.
The data are typically sparse, thus the number of attributes per object
is bounded as follows: 13 ≤ |g 0 | ≤ 947. For objects per attribute we have
18 ≤ |m0 | ≤ 159. From this context we computed formal concepts of the form
(advertisers, bids) that represent market sectors. Formal concepts of this form
can be further used for recommendation to the companies on the market, which
did not buy bids contained in the intent of the concept.
This can also be represented as association rules of the form “If an advertiser
bought bid a, then this advertiser may buy term b” See [202] for the use of
association rules in recommendation systems.
To make recommendations we used the following steps:

1. D-miner algorithm for detecting large market sectors as concepts and our
biclustering algorithm;
2. Coron system for constructing association rules;
3. Construction of association metarules using morphological analysis;
4. Construction of association metarules using ontologies (thematic catalogs).

Detecting large market sectors with D-miner and OA-biclustering. The D-miner
algorithm [203] constructs the set of concepts satisfying given constraints on sizes
of extents and intents (i.e. intersection of icebergs and dual icebergs). D-miner
takes as input a context and two parameters: minimal admissible extent and
intent sizes and outputs a “band” of the concept lattice: all concepts satisfying
constraints given by parameter values (|intent| ≥ m and |extent| ≥ n, where
m, n ∈ N, see table 6).

Example 20. We provide examples of two intents of formal concepts for the case
|L| = 53, where |L| is a number of formal concepts obtained by D-miner.
Hotel market.
{ angeles hotel los, atlanta hotel, baltimore hotel, dallas hotel, denver hotel, hotel
chicago, diego hotel san, francisco hotel san, hotel houston, hotel miami, hotel new or-
leans, hotel new york, hotel orlando, hotel philadelphia, hotel seattle, hotel vancouver}
Weight loss drug market.
66 Dmitry I. Ignatov

Table 6. D-miner results.

Minimal extent Minimal intent Number of

size size concepts
0 0 8 950 740
10 10 3 030 335
15 10 759 963
15 15 150 983
15 20 14 226
20 15 661
20 16 53
20 20 0

{ adipex buy, adipex online, adipex order, adipex prescription, buy didrex,
buy ionamin, ionamin purchase, buy phentermine, didrex online, ionamin on-
line, ionamin order, online order phentermine, online phentermine, order phen-
termine, phentermine prescription, phentermine purchase } 
Applying the biclustering algorithm to our data we obtained 87 OA-biclusters
(ρ = 0.85), which is much less than the number of concepts found by D-miner.
Expert interpretation of these biclusters implies that each market described by
formal concepts found by D-miner (where each market can be represented by
several formal concepts) corresponds to a bicluster among these 87. The number
of formal concepts generated by D-miner becomes feasible for human interpreta-
tion if there are no more than 20 firms and about 15 terms. For these thresholds
D-miner could find only large markets and ignored important average-size mar-
kets. For our data these ignored markets were, e.g. car and flower markets, which
were found using biclustering approach.
Example 21. Flower market OA-bicluster.
({ 24, 130, 170, 260, 344, 415, 530, 614, 616, 867, 926, 1017, 1153, 1160, 1220, 1361,
1410, 1538, 1756, 1893 }, {’anniversary flower’, ’arrangement flower’, ’birthday flower’,
’bouquet flower’, ’buy flower’, ’buy flower online’, ’delivery flower’, ’flower fresh’, ’flower
gift’, ’flower line’, ’flower online’, ’flower online order’, ’flower online send’, ’flower online
shop’, ’flower rose’, ’flower send’, ’flower shop’, ’flower sympathy’, ’red rose’}), with
ρ ≈ 0.84 

Recommendations based on association rules. Using the Coron system (see [204])
we construct the informative basis of association rules [205].
Example 22. Here are some examples of association rules:
– {evitamin} → {cvitamin}, supp=31 [1.55%] and conf=0.86;
– {gif t graduation} → {anniversary gif t}, supp=41 [2.05%] and conf=0.82.

 Introduction to Formal Concept Analysis and Its Applications in IR 67

The value supp = 31 of the first rule means that 31 companies bought phrases
“e vitamin” and “c vitamin”. The value conf = 0.861 means that 86,1% com-
panies that bought the phrase “e vitamin” also bought the phrase “c vitamin”.
To make recommendations for each particular company one may use an
approach proposed in [202]. For company f we find all association rules, the
antecedent of which contain all the phrases bought by the company, then we
construct the set Tu of unique advertising phrases not bought by the company
f before. Then we order these phrases by decreasing of confidence of the rules
where the phrases occur in the consequences. If buying a phrase is predicted by
multiple rules, we take the largest confidence.

Morphology-based Metarules Each attribute of our context is either a word or

a phrase. Obviously, synonymous phrases are related to same market sectors.
The advertisers companies have usually thematic catalogs composed by experts,
however due to the huge number of advertising terms manual composition of cat-
alogs is a difficult task. Here we propose a morphological approach for detecting
similar terms.
Let t be an advertising phrase consisting of several words (here we disregard
the word sequence): t = {w1 , w2 , . . . , wn }. A stem is the root or roots of a word,
together with any derivational affixes, to which inflectional affixes are added
[206]. The stem of word wi is denoted by si = stem(w S i ) and the set of stems
of words of the phrase t is denoted by stem(t) = stem(wi ), where wi ∈ t.
i
Consider the formal context KT S = (T, S, IT S ), where T isSthe set of all phrases
and S is the set of all stems of phrases from T , i.e. S = stem(ti ). Then tIs
i
denotes that the set of stems of phrase t contains s.
In this context we construct rules of the form t → sIi T S for all t ∈ T , where
Its
(.) denotes the prime operator in the context KT S . Then the a morphology-
based metarules of the context KT S (we call it a metarule, because it is not based
on experimental data, but on implicit knowledge resided in natural language
FT
constructions) corresponds to t −−→ sIi T S , an association rule of the context
KF T = (F, T, IF T ). If the values of support and confidence of this rule in context
KF T do not exceed certain thresholds, then the association rules constructed
from the context KF T are considered not very interesting.

Example 23. An example of an input context for morphological association rules.



Metarules of the following forms seem also to be reasonable. First, one can
F T S IT S
look for rules of the form t −−→ si , i.e., rules, the consequent of which
i
contain all terms containing at least one word with the stem common to a word
in the antecedent term. Obviously, constructing rules of this type may result in
the fusion of phrases related to different market sectors, e.g. “black jack” and
FT
“black coat”. Second, we considered rules of the form t −−→ ( si )IT S , i.e., rules
S
i
with the consequent with the set of stems being the same as the set of stems
68 Dmitry I. Ignatov

Table 7. A toy example of context KF T for “long distance calling” market.

firm \ phrase call calling calling carrier cheap

distance distance distance distance distance
long long long plan long long
f1 x x x
f2 x x x
f3 x x
f4 x x x
f5 x x x x

Table 8. A toy example of context KT S for “long distance calling” market.

phrase \ stem call carrier cheap distanc long plan

call distance long x x x
calling distance long x x x
calling distance long plan x x x x
carrier distance long x x x
cheap distance long x x x

of the antecedent. Third, we also propose to consider metarules of the form

FT
t1 −−→ t2 , where tI2T S ⊆ tI1T S . These are rules with the consequent being sets of
stems that contain the set of stems of the antecedent.
Example 24. An example of metarules.
FT
– t −−→ sIi T S
{last minute vacation} → {last minute travel}
supp= 19 conf= 0,90
FT S
– t −−→ sIi T S
i
{mail order phentermine} → {adipex online order, . . . ,
phentermine purchase, phentermine sale}
supp= 19 conf= 0,95
FT
– t −−→ ( si )IT S
S
i
{distance long phone} → {call distance long phone, . . . ,
distance long phone rate, distance long phone service}
supp= 37 conf= 0,88
FT
– t1 −−→ t2 , tI2T S ⊆ tI1T S
{ink jet} → {ink}, supp= 14 conf= 0,7


 Introduction to Formal Concept Analysis and Its Applications in IR 69

Experimental Validation For validation of association rules and metarules we

used an adapted version of cross-validation. The training set was randomly di-
vided into 10 parts, 9 of which were taken as the training set and the remaining
part was used as a test set. The confidence of rules averaged over the test set is
almost the same as the min conf for the training set, i.e., (0.9−0.87)/0.9 ≈ 0.03.
Note that the use of morphology is completely automated and allows one to
find highly plausible metarules without data on purchases. The rules with low
support and confidence may be tested against recommendation systems such as
Google AdWords, which uses the frequency of queries for synonyms. Thus 90%
of recommendations (words) for ontological rules (see [96]) were contained in the
list of synonyms output by AdWords.

6.6 FCA-based recommender systems

Motivated by prospective applications of Boolean Matrix Factorisation (BMF) in
the context of Recommender Systems (RS) we proposed an FCA-based approach
which follows user-based k-nearest neighbours strategy [94]. Another approach
similar to MF is biclustering, which has also been successfully applied in recom-
mender system domain [207,96]. As we have mentioned, FCA can be also used
as a biclustering technique and there are several examples of its applications in
the recommender systems domain [90,92]. A parameter-free approach that ex-
ploits a neighbourhood of the object concept for a particular user also proved
its effectiveness [208].
Belowe we discuss our recent studies in application of BMF for RS. In the
recommender systems domain, the context is any auxiliary information concern-
ing users (like gender, age, occupation, living place) and/or items (like genre of
a movie, book or music), which shows not only a user’s mark given to an item
but explicitly or implicitly describes the circumstances of such evaluation (e.g.,
including time and place) [209].
From representational viewpoint an auxiliary information can be described by
a binary relation, which shows that a user or an item posses a certain attribute-
value pair.
As a result one may obtain a block matrix:
 
R Cuser
I= ,
Citem O
where R is a utility matrix of users’ ratings to items, Cuser represents context
information of users, Citem contains context iformation of items and O is a zero-
filled matrix.
Example 25. An example of a rating matrix enriched by user-feature and item-
feature auxiliary information.

Example 26. In case of more complex rating’s scale the ratings can be reduced
to binary scale (e.g., “like/dislike”) by binary thresholding or by FCA-based
scaling.
70 Dmitry I. Ignatov

Table 9. Adding auxiliary information

Movies Gender Age

Brave Termi- Gladi- Slum- Hot God- M F 0-20 21-45 46+
Heart nator ator dog Snow father
Million-
aire
Anna 5 5 5 2 + +
Vladimir 5 5 3 5 + +
Katja 4 4 5 4 + +
Mikhail 3 5 5 5 + +
Nikolay 2 5 4 + +
Olga 5 3 4 5 + +
Petr 5 4 5 4 + +
Drama + + + + +
Action + + + +
Comedy + +

Table 10. Derived Boolean utility matrix enriched by auxiliary information

m1 m2 m3 m4 m5 m6 f1 f2 f3 f4 f5
u1 1 0 1 1 0 0 0 1 1 0 0
u2 1 0 1 1 0 0 1 0 0 1 0
u3 1 0 1 1 0 1 0 1 0 1 0
u4 1 0 1 1 0 0 1 0 0 1 0
u5 0 0 0 0 1 1 1 0 0 0 1
u6 1 0 1 1 0 0 0 1 1 0 0
u7 1 0 0 1 1 1 1 0 0 0 1
g1 1 0 1 1 1 1 0 0 0 0 0
g2 0 1 1 0 1 1 0 0 0 0 0
g3 1 0 0 1 0 0 0 0 0 0 0

Once a matrix of ratings is factorized we need to learn how to compute

recommendations for users and to evaluate whether a particular method handles
this task well.
Given the factorized matrices already well-known algorithm based on the
similarity of users can be applied, where for finding k nearest neighbors we use
not the original matrix of ratings R ∈ Rm×n , but the matrix I ∈ Rm×f , where m
is a number of users, and f is a number of factors. After the selection of k users,
which are the most similar to a given user, based on the factors that are peculiar
to them, it is possible, based on collaborative filtering formulas to calculate the
prospective ratings for a given user.
 Introduction to Formal Concept Analysis and Its Applications in IR 71

After generation of recommendations the performance of the recommender

system can be estimated by such measures as Mean Absolute Error (MAE),
Precision and Recall.
Collaborative recommender systems try to predict the utility (in our case
ratings) of items for a particular user based on the items previously rated by
other users.
Memory-based algorithms make rating predictions based on the entire col-
lection of previously rated items by the users. That is, the value of the unknown
rating ru,m for a user u and item m is usually computed as an aggregate of the
ratings of some other (usually, the k most similar) users for the same item m:

ru,m = aggrũ∈Ũ rũ,m ,

where Ũ denotes a set of k users that are the most similar to user u, who
have rated item m. For example, the function aggr may have the following form
[209]:

P
sim(ũ, u) · rũ,m ,
ũ∈Ũ
ru,m = P .
sim(u, ũ)
ũ∈Ũ

Similarity measure between users u and ũ, sim(ũ, u), is essentially an inverse
distance measure and is used as a weight, i.e., the more similar users c and ũ
are, the more weight rating rũ,m will carry in the prediction of rũ,m .
Similarity between two users is based on their ratings of items that both
users have rated. There are several popular approaches: Pearson correlation,
cosine-based, and Hamming-based similarities.
We mainly use the cosine-based and normalised Hamming-based similarities.
To apply this approach in case of FCA-based BMF recommender algorithm
we simply consider the user-factor matrices obtained after factorisation of the
initial data as an input.

Example 27. For the input matrix in Table 10 one can find the following covering
factors:

({u1 , u3 , u6 , u7 , g1 , g2 }, {m1 , m4 }), ({u2 , u4 }, {m2 , m3 , m6 , f1 , f4 }),

({u5 , u7 }, {m5 , m6 , f1 , f5 }), ({u1 , u6 }, {m1 , m3 , m4 , f2 , f3 }),
({u5 , u7 , g1 , g3 }, {m5 , m6 }), ({u2 , u3 , u4 }, {m3 , m6 , f4 }),
({u2 , u4 , g3 }, {m2 , m3 , m6 }), ({u1 , u3 , u6 , g1 }, {m1 , m3 , m4 }),
({u1 , u3 , u6 }, {m1 , m3 , m4 , f2 }).

The corresponding decomposition is below:

72 Dmitry I. Ignatov

 
100100011  
0 1 0 0 0 1 1 0 0 10010000000
1 0 0 0 0 1 0 1 1 0 1 1 0 0 1 1 0 0 1 0
   
  0 0 0 0 1 1 1 0 0 0 1
0 1 0 0 0 1 1 0 0  
0 0 1 0 1 0 0 0 0 1 0 1 1 0 0 0 1 1 0 0
   
  ◦ 0 0 0 0 1 1 0 0 0 0 0
1 0 0 1 0 0 0 1 1  
1 0 1 0 1 0 0 0 0 0 0 1 0 0 1 0 0 0 1 0
   
  0 1 1 0 0 1 0 0 0 0 0
1 0 0 0 1 0 0 1 0  
  1 0 1 1 0 0 0 0 0 0 0
0 0 0 0 1 0 1 0 0
10110001000
100000000


However, in this case in the obtained user profiles vectors most of the com-
ponents are getting zeros, and thus we lose similarity information.
To smooth the loss effects we proposed the following weighted projection:
P
Iuv · Qf v
Iu· · Qf · v∈V
P̃uf = = P ,
||Qf · ||1 Qf v
v∈V

where P˜uf indicates whether factor f covers user u, Iu· is a binary vector
describing profile of user u, Qf · is a binary vector of items belonging to factor f
(the corresponding row of Q in decomposition eq. (3)). The coordinates of the
obtained projection vector lie within [0; 1].

Example 28. For Table 9 the weighted projection is as follows:

 1
1 5 0 1 0 31 13 1 1

0 1 1 1 1 1 1 1 1 
 3 21 45 21 2 3 4 
1
 5 41 15 21 1 3 11 11 

0 1
 2 2 5 2 12 11 3 4 

0 1 0 1
3 3 0 0

P̃ =  5
1 1 0 1 0 1 1 1 1 .
 25 1 31 13 2 1 
1 1 1 
 25 1 25 32 23 3 23 
1
 25 21 15 1 32 3 11 41 

0
5 2 5 1 3 1 3 4

1 2 1
10 0 5 0 0 0 3 2

The proposed approach and compared ones have been implemented in C++35
and evaluated on MovieLens-100k data set. This data set features 100000 ratings
in five-star scale, 1682 Movies, Contextual information about movies (19 genres),
943 users (each user has rated at least 20 movies), and demographic info for the
35
https://github.com/MaratAkhmatnurov/BMFCARS
 Introduction to Formal Concept Analysis and Its Applications in IR 73

users (gender, age, occupation, zip (ignored)). The users have been divided into
seven age groups: under 18, 18-25, 26-35, 36-45, 45-49, 50-55, 56+.
Five star ratings are converted to binary scale by the following rule:
(
1, Rij > 3,
Iij =
0, else

The scaled dataset is split into two sets according to bimodal cross-validation
scheme [210]: the training set and the test set with a ratio 80:20, and 20% of
ratings in the test set are hidden.

0.4 1

0.35 0.8

Precision
MAE

0.3 0.6

0.25 0.4

0.2 0.2
0 20 40 60 80 100 0 20 40 60 80 100
Number of neighbours Number of neighbours

0.4 0.4
BMF80+context
0.3 0.3 SVD85+context
F−measure

SVD85
Recall

0.2 0.2

0.1 0.1

0 0
0 20 40 60 80 100 0 20 40 60 80 100
Number of neighbours Number of neighbours

Fig. 25. Different approaches of matrix factorisation

We found out that MAE of our BMF-based approach is significantly lower

than MAE of SVD-based approach for almost the same number of factor at fixed
coverage level of BMF and p-level of SVD. The Precision of BMF-based approach
is slightly lower when the number of neighbours is about a couple of dozens and
comparable for the remaining part of the observed range. The Recall is lower
that results in lower F-measure. It can be explained by different nature of factors
in these factorisation models. The proposed weighted projection alleviates the
information loss of original Boolean projection resulting in a substantial quality
gain. We also revealed that the presence of auxiliary information results in a
small quality increase (about 1-2%) in terms of MAE, Recall and Precision.
74 Dmitry I. Ignatov

In our previous study, with the original BMF-based scheme (weighting is not
used), we obtained comparable results in terms of MAE, and both Precision and
Recall [94,145].

6.7 FCA-based approach for near-duplicate documents detection

From the dawn of web search engines, the problem of finding near-duplicate
documents in the web search results is crucial for providing users with relevant
documents [211,212,213].
Below we shortly describe our studies on near duplicate detection [81] within
a competition “Internet mathematics” organised by Yandex and ROMIP (Rus-
sian Information Retrieval Evaluation Seminar) in 2004–2005: our project “Op-
timisation of search for near duplicates in the web: images and similarity” was
selected, as the rest 33 projects, out of 252 applications 36 .
As experimental data the ROMIP collection of web documents from narod.ru
domain37 was provided; it consists of 52 files of general size 4.04 GB. These files
contained 530 000 web pages from narod.ru domain. Each document from the
collection has size greater or equal to 10 words. For experiments the collection
was partitioned into several parts consisting of three to 24 such files (from 5% to
50% percent of the whole collection). As an evaluation benchmark for recall and
precision calculation we use the list of duplicate pairs provided by Yandex; the
duplicate pairs were identified for all document pairs via Perl String::Similarity
with 85% similarity threshold.
For composing document images we followed a popular shingling approach
[212]. For each text, the program shingle with two parameters (length and
offset) generate contiguous subsequences of size length such that the distance
between the beginnings of two subsequent substrings is offset. The set of se-
quences obtained in this way is hashed so that each sequence receives its own
hash code. From the set of hash codes that corresponds to the document a fixed
size (given by parameter) subset is chosen by means of random permutations
described in [212,214,215]. The probability of the fact that minimal elements in
permutations on hash code sets of shingles of documents A and B (these sets are
denoted by FA and FB , respectively) coincide, equals to the similarity measure
of these documents sim(A, B):

|FA ∩ FB |
sim(A, B) = P [min{π(FA )} = min{π(FB )}] = .
|FA ∪ FB |
Further we used FCA to define cluster of near duplicate documents.
Let KDF = (D, F, I ⊆ D × F ) be a context of documents, where D is
a set of documents, F is a set of hash codes (fingerprints), and I shows that a
document d has an attribute f whenever dIf .
For a subset of documents A ⊆ D, A0 describe their similarity in terms of
common fingerprints, and the closed set A00 is a cluster of similar documents.
36
https://academy.yandex.ru/events/imat/grant2005/
37
http://romip.ru/en/collections/narod.html
 Introduction to Formal Concept Analysis and Its Applications in IR 75

To find all near duplicate clusters, we need to enumerate all intents of the
context KF D = (F, D, I ⊆ F × D) such that their common set of fingerprints
exceeds a threshold set by user.
In fact, to this end we need to use nothing but frequent itemsets of documents.
A set of documents A ⊆ D is called k-frequent if |A0 | > k, where k is a
parameter.

Program Implementation Software for experiments with syntactical representa-

tion comprise the units that perform the following operations:

1. XML Parser (provided by Yandex): it parses XML packed collections of web

documents
2. Removing html-markup of the documents
3. Generating shingles with given parameters length-of-shingle, offset
4. Hashing shingles
5. Composition of document image by selecting subsets (of hash codes) of shin-
gles by means of methods n minimal elements in a permutation and minimal
elements in n permutations.
6. Composition of the inverted table the list of identifiers of documents shingle
thus preparing data to the format of programs for computing closed itemsets.
7. Computation of clusters of k-similar documents with FPmax* algorithm:
the output consists of strings, where the first elements are names (ids) of
documents and the last element is the number of common shingles for these
documents.
8. Comparing results with the existing list of duplicates (in our experiments
with the ROMIP collection of web documents, we were supplied by a pre-
computed list of duplicate pairs).

This unit outputs five values: 1) the number of duplicate pairs in the ROMIP
collection, 2) the number of duplicate pairs for our realisation, 3) the number
of unique duplicate pairs in the ROMIP collection, 4) the number of unique
duplicate pairs in our results, 5) the number of common pairs for the ROMIP
collection and our results.
In step 7, we used a leader in time efficiency, the algorithm FPmax* [216],
from the competition organised in series of workshops on Frequent Itemset Min-
ing Implementations (FIMI)38 .

Experimental results In our experiments we used Cluto 39 , a software package for

clustering high-dimensional datasets including those from information retrieval
domain, for comparison purposes. We chose the repeated-bisecting algorithm
that uses the cosine similarity function with a 10-way partitioning (ClusterRB),
which is mostly scalable according to its author [217]. The number of clusters was
a parameter, documents were given by sets of attributes, fingerprints in our case.
The algorithm outputs a set of disjoint clusters. Algorithms from FIMI repository
38
http://fimi.ua.ac.be/
39
http://glaros.dtc.umn.edu/gkhome/views/cluto
76 Dmitry I. Ignatov

can process very large datasets, however, to compare with Cluto (which is much
more time consuming as we show below) we took collection narod.1.xml that
contains 6941 documents.
Shingling parameters used in experiments were as follows: the number of
words in shingles was 10 and 20, the offset was always taken to be 1 (which means
that the initial set of shingles contained all possible contiguous word sequences
of a given length). The sizes of resulting document images were taken in the
interval 100 to 200 shingles. As frequency thresholds defining frequent closed sets
(i.e., the numbers of common shingles in document images from one cluster) we
experimentally studied different values in intervals, where the maximal value
is equal to the number of shingles in the document image. For example, the
interval [85, 100] for document images with 100 shingles, the interval [135, 150]
for document images of size 150, etc. Obviously, choosing the maximal value of
an interval, we obtain clusters where document images coincide completely.
For parameters taking values in these intervals we studied the relation be-
tween resulting clusters of duplicates and ROMIP collection of duplicates, which
consists of pairs of web documents that are considered to be near duplicates. Sim-
ilarity of each pair of documents in this list is based on Edit Distance measure,
two documents were taken to be duplicates by authors of this testbed if the value
of the Edit Distance measure exceeds threshold 0.85. As we show below, this def-
inition of a duplicate is prone to errors, however making a testbed by manual
marking duplication in a large web document collection is hardly feasible. Unfor-
tunately, standard lists of near-duplicates were missing at that period even for
standard corpora such as TREC or Reuters collection [218]. For validating their
methods, researchers create ad-hoc lists of duplicates using slightly transformed
documents from standard collections. Now the situation is drastically better, see,
for example, workshop series on Plagiarism Analysis, Authorship Identification,
and Near-Duplicate Detection (PAN)40 .
In our study for each such pair we found an intent that contains both el-
ements of the pair, and vice versa, for each cluster of very similar documents
(i.e., for each corresponding closed set of documents with more than k common
description units) we take each pair of documents in the cluster and looked for
the corresponding pair in the ROMIP collection. As result we obtain the number
of common number of near duplicate pairs found by our method and those in the
ROMIP collection, and the number of unique pairs of HSE duplicates (document
pairs occurring in a cluster of “very similar documents” and not occurring in
the ROMIP collection). The results of our experiments showed that the ROMIP
collection of duplicates, considered to be a benchmark, is far from being perfect.
First, we detected that a large number of false duplicate pairs in this collection
due to similar framing of documents. For example the pages with the following
information in table 11 about historical personalities 1 and 2 were declared to
be near duplicates.

40
http://www.uni-weimar.de/medien/webis/events/pan-15/pan15-web/
plagiarism-detection.html
 Introduction to Formal Concept Analysis and Its Applications in IR 77

Table 11. Information about historical personalities

1. Garibald II, Duke of Bavaria 2. Giovanni, Duke of Milan

Short information: Short information:
Full Name: Garibald Full Name: Giovanni Visconti
Date of birth: unknown Date of birth: unknown
Place of birth: unknown Place of birth: unknown
Date of death: 610 Date of death: 1354
Place of death: unknown Place of death: unknown
Father: Tassilo I Duke of Bavaria Father: Visconti Matteo I,
the Great Lord of Milan
Mother: uknown Mother: uknown

However these pages, as well as many other analogous false duplicate pairs in
ROMIP collection do not belong to concept-based (maximal frequent) clusters
generated in our approach.
Second, in our study we also looked for false duplicate clusters in the ROMIP
collection, caused by transitive closure of the binary relation “X is a duplicate
of Y ” (as in the typical definition of a document cluster in [215]). Since the
similarity relation is generally not transitive, the clusters formed by transitive
closure of the relation may contain absolutely dissimilar documents. Note that if
clusters are defined via maximal frequent itemsets (subsets of attributes) there
cannot be effects like this, because documents in these clusters share necessarily
large itemsets (common subsets of attributes).
We analysed about 10000 duplicate document pairs and found four rather
big false duplicate clusters about 50-60 documents each. Further studies on this
collection see in [219].
We shortly summarise experimental results below:

– FPmax* (F-measure=0.61 and elapsed time 0.6 seconds), ClusterRB (F-

measure=0.63 and elapsed time 4 hours);
– For FPMax* the number of single document cluster is 566, for ClusterRB
4227;
– The total number of clusters for FPmax* is 903 versus 5000 for Cluto 903;
– The number of NDD clusters for FPmax* is 337 versus 773 Cluto.

Graphs and tables show that for 5000 clusters the output of ClusterRB has al-
most the same value of F-measure (0.63) as FPmax* for threshold 150 (F1=0,61).
However, computations took 4 hours for ClusterRB and half a second for FP-
max*.
We continued our developments of that time and developed GUI and du-
plicate document generator (for a provided text collection) for testing purposes
[220].The archive of these projects is freely available at Bitbucket41 .
Later on, we proposed a prototype of near-duplicate detection system for web-
shop owners. It’s a typical situation for this online businesses to buy description
41
https://bitbucket.org/dimanomachine/nearduplicatesarch
78 Dmitry I. Ignatov

Table 12. Comparison of the obtained clusters in terms of pairs of near duplicate
documents

The number of ROMIP duplicates: 2997

The number of NDD found by FPmax*: 2722
The number of NDD found by Cluto: 2897
The number of unique NND pairs of ROMIP: 1155
The number of unique NDD pairs found by FPmax*: 1001
The number of unique NDD pairs found by Cluto: 1055
The number of common NDD pairs for FPmax* and ROMIP: 1721
The number of common NDD pairs for Cluto and ROMIP: 1842

of their goods from so-called copyrighters. A copyrighter may cheat from time to
time and provide the owner with some almost identical descriptions for different
items. In that study we demonstrated how we can use FCA for revealing and fast
online clustering of such duplicates in a real online perfume shop. Our results
were also applicable for near duplicate detection in collections of R&D project’s
documents [221].

6.8 Triadic FCA for IR-tasks in Folksonomies

Four our data mining studies on triclustering (see Section 5.1 and [109,112,111])
folksonomic data became a shootingrange since the first efficient FCA-based
algorithm to mine tiradic data was proposed for mining communities in folk-
sonomies [45].
But it is a rich field with interactive resource-sharing systems like Bibson-
omy42 , CiteULike43 , Flickr44 and Delicious45 that need fully-fledged IR func-
tionality including retrieval, ranking and recommendations. For example, Bib-
sonomy is a social bookmarking and publication management system. Apart
from DBLP46 [222] that only collects, stores bibliographic data and provide
publication and author search, Bibsonomy allows to create user’s own lists of
bibliographic bookmarks, use tags and social interactions.
As we have mentioned in Section 5.1, the underlying Folksonomy structure
is a formal tricontext K = (U, T, R, Y ) with U being a set of users, T a set tags,
and R a set of resources, where Y ⊆ U × T × R relates entities from these three
sets. Sometimes, a user-specific subtag/supertag-relation  is also included into
the definition, i.e. ⊆ U × T × T .
We shortly discuss main IR tasks that folksonomic data give rise.
First of all, we have to say that traditional PageRank cannot be directly
applied to folksonomies. The authors of paper [161] modified the PageRank
42
http://www.bibsonomy.org/
43
http://www.citeulike.org/
44
https://www.flickr.com/
45
https://delicious.com/
46
http://dblp.uni-trier.de/
 Introduction to Formal Concept Analysis and Its Applications in IR 79

algorithm for folksonomic data by considering the input triadic data as an undi-
rected tripartite graph. The weights for each type of edge were assigned accord-
ing to the occurrences of the third entity, e.g. an edge u, t being weighted with
|r ∈ R : (u, t, r) ∈ Y |, the number occurrences of the related tags.
Formally, the weight spreading condition looks as follows:

w ← αw + βAw + γp, where

A is the row-stochastic version of the graph adjacency, p is a preference vector,
α, β, γ ∈ [0, 1] are initial parameters with α + β + γ = 1. Thus, α regulates the
speed of convergence, while the proportion between β and γ controls the influence
of the preference vector.
However, the first results on Delicious data were rather discouraging even
with a term-frequency ranker combination, the resulting ranking was similar
(though not identical) to the the initial edge weights. It resulted in authors’
own ranking algorithm FolkRank, which takes into account the difference in the
resulting rankings with and without preference vector [161].
In that paper the authors formulated peculiar tasks:

– Documents that are of potential interest to a user can be suggested to him.

– Other related tags can be suggested to a user. Thus, FolkRank addition-
ally considers the tagging behavior of other users and can be used for tag
recommendations.
– Other users that work on related topics can be made explicit and this facil-
itates knowledge transfer and formation of user communities.

Later on, they implemented (not only) all this features in the Bibsonomy
systems [97].
Moreover, during those studies they admitted that search query logs natu-
rally forms folksonomic data, (users, queries, resources), where the resources
are those that were clicked by a user after performing a query [223]. Predictably,
they gave a name logsonomy to this new data structure. When Bibsonomy was
at the early stages, it faced spam abuse problem and in 2008 ECML PKDD
discovery challenge 47 addressed this problem. The year after the challenging
problem 48 were recommendations for Bibsonomy and it resulted in new fruitful
algorithms [224].

6.9 Exploring taxonomies of web site users

In 2006 we participated in analysis of web sites audience in collaboration with

SpyLog company (now OpenStat49 )[225].
Owners of a web-site are often interested in analysing groups of users of
their site. Information on these groups can help to optimise the structure and
47
http://www.kde.cs.uni-kassel.de/ws/rsdc08/
48
http://www.kde.cs.uni-kassel.de/ws/dc09/
49
https://www.openstat.com/
80 Dmitry I. Ignatov

Fig. 26. An example of Bibsonomy interface

contents of the site. For example, interaction with members of each group may
be organized in a special manner. In the performed study we used an approach
based on formal concepts for constructing taxonomies of groups of web users.
For our experiments we have chosen four target websites: the site of the State
University Higher School of Economics, an e-shop of household equipment, the
site of a large bank, and the site of a car e-shop (the names of the last three
sites cannot be disclosed due to legal agreements).
Users of these sites are described by attributes that correspond to other sites,
either external (from three groups of sites: finance, media, education) or internal
(web-pages of the site). More precisely, initial “external” data consists of user
records each containing the user id, the time when the user first entered this site,
the time of his/her last visit, and the total number of sessions during the period
under consideration. An “internal” user record, on the other hand, is simply a
list of pages within the target website visited by a particular user.
By “external” and “internal” taxonomies we mean (parts of) concept lattices
for contexts with either “external” or “internal” attributes. For example, the
external context has the form Ke = (U, Se , Ie ), where U is the set of all users
of the target site, Se is the set of all sites from a sample (not including the
target one), the incidence relation Ie is given by all pairs (u, s) : u ∈ U, s ∈ Se ,
 Introduction to Formal Concept Analysis and Its Applications in IR 81

such that user u visited site s. Analogously, the internal context is of the form
Ki = (U, Si , Ii ), where Si is the set of all own pages of the target site.
A concept of this context is a pair (A, B) such that A is a group of users that
visited together all other sites from B.
As we have mentioned, one of the target websites was the site of our univer-
sity50 .
We received “external” data with the following fields for each user-site pair:
(user id, time of the first visit, time of the last visit, total number of
sessions during the period). “Internal” data have almost the same format
with an additional field url page, which corresponds to a particular visited page
of the target site.
The provided information was gathered from about 10000 sites of Russian
segment of Internet (domain .ru). Describing users in terms of sites they vis-
ited, we had to tackle the problem of dimensionality, since the resulting concept
lattices can be very large (exponential in the worst case in terms of objects or
attributes). To reduce the size of input data we used the following techniques.
For each user we selected only those sites that were visited by more than
a certain number of times during the observation period. This gave us infor-
mation about permanent interests of particular users. Each target web site was
considered in terms of sites of three groups: newspaper sites, financial sites, and
educational sites.
However, even for large reduction of input size, concept lattices can be very
large. For example, a context of size 4125 × 225 gave rise to a lattice with 57 329
concepts.
To choose interesting groups of users we employed stability index of a concept
defined in [226,227] and considered in [88] (in slightly different form) as a tool for
constructing taxonomies. On one hand, stability index shows the independence
of an intent on particular objects of extent (which may appear or not appear in
the context depending on random factors). On the other hand, stability index of
a concept shows how much extent of a concept is different from similar smaller
extents (if this difference is very small, then its doubtful that extent refers to a
“stable category”). For detailed motivation of stability indices see [226,227,88].

Definition 33. Let K = (G, M, I) be a formal context and (A, B) be a formal

concept of K. The stability index σ of (A, B) is defined as follows:

|{C ⊆ A|C 0 = B}|

σ(A, B) = .
2|A|

Obviously, 0 ≤ σ(A, B) ≤ 1.
The stability index of a concept indicates how much the concept intent de-
pends on particular objects of the extent. A stable intent (with stability index
50
www.hse.ru
82 Dmitry I. Ignatov

close to 1) is probably “real” even if the description of some objects is “noisy”.

In application to our data, the stability index shows how likely we are to still
observe a common group of interests if we ignore several users. Apart from being
noise-resistance, a stable group does not collapse (e.g., merge with a different
group, split into several independent subgroups) when a few members of the
group stop attending the target sites.
In our experiments we used ConceptExplorer for computing and visualising
lattices and their parts.
We compared results of taking most stable concepts (with stability index ex-
ceeding a threshold) with taking an “iceberg” lattice. The results look correlated,
but nevertheless, substantially different. The set of stable extents contained very
important, but not large groups of users.
In Figs.12,27 we present parts of a concept lattice for the HSE web site
described by “external” attributes which were taken to be Russian e-newspapers
visited by users of www.hse.ru during one month (September 2006) more than 20
times. Fig. 12 presents an iceberg with 25 concepts having largest extent. Many
of the concepts correspond to newspapers that are in the middle of political
spectrum, read “by everybody” and thus, not very interesting in characterising
social groups.
Fig. 27 presents an ordered set of 25 concepts having largest stability index.
As compared to the iceberg, this part of the concept lattice contains several
sociologically important groups such as readers of “ExpressGazeta” (“yellow
press”), Cosmopolitan, Expert (high professional analytical surveys) etc.

Fig. 27. Ordered set of 25 concepts with largest stability

 Introduction to Formal Concept Analysis and Its Applications in IR 83

6.10 FCA-based models for crowdsourcing

The success of modern collaborative technologies is marked by the appearance

of many novel platforms for holding distributed brainstorming or carrying out
so called “public examination”. There are a lot of such crowdsourcing companies
in the USA (Spigit 51 , BrightIdea 52 , InnoCentive 53 etc.) and Europe (Imag-
inatik 54 ). There is also the Kaggle platform 55 which is the most beneficial
for data practitioners and companies that want to select the best solutions for
their data mining problems. In 2011 Russian companies launched business in
that area as well. The two most representative examples of such Russian compa-
nies are Witology 56 and Wikivote 57 . Several all-Russian projects have already
been finished successfully (for example, Sberbank-2158 , National Entrepreneurial
Initiative 59 etc.). The core of such crowdsourcing systems is a socio-semantic
network [228,229], which data requires new approaches to analyze. Before we
tried to accommodate FCA as a methodological base for the analysis of data
generated by such collaborative systems [164].
As a rule, while participating in a project, users of such crowdsourcing plat-
forms [230] discuss and solve one common problem, propose their ideas and
evaluate ideas of each other as experts. Finally, as a result of the discussion and
ranking of users and their ideas we get the best ideas and users (their generators).
For deeper understanding of users’s behavior, developing adequate ranking cri-
teria and performing complex dynamic and statistic analyses, special means are
needed. Traditional methods of clustering, community detection and text mining
need to be adapted or even fully redesigned. Earlier we described models of data
used in crowdsourcing projects in terms of FCA. Furthermore, we presented the
collaborative platform data analysis system CrowDM (Crowd Data Mining), its
architecture and methods underlying the key steps of data analysis [164].
The principles of these platforms’ work are different from the work of online-
shops or specialized music/films recommender websites. Crowdsourcing projects
consist of several stages and results of each stage substantially depend on the pre-
vious stage results. That’s why the existing models of the recommender systems
should be adapted properly. In the accompanion paper [231] or in its shorter
predecessors [232,233], we present new methods for making recommendations
based on FCA and OA-biclustering (see Section 5.1): The original methods of
idea recommendation (for voting stage), like-minded persons recommendation
(for collaboration) and antagonists recommendation (for counteridea generation

51
http://spigit.com/
52
http://www.brightidea.com/
53
http://www.innocentive.com/
54
http://www.imaginatik.com/
55
http://www.kaggle.com
56
http://witology.com/
57
http://www.wikivote.ru/
58
http://sberbank21.ru/
59
http://witology.com/en/clients_n_projects/3693/
84 Dmitry I. Ignatov

stage). The last recommendation type is very important for stimulating user’s
activity on Witology platform during the stage of counteridea generation.

7 FCA in Ontology Modeling and Attribute Exploration

Applications of FCA in ontology modeling and its relations with Semantic Web
deserve a special treatment. However, we shortly mention several interesting
approaches and showcase an interactive technique which can be used for ontology
and knowledge bases refinement and building.

– Attribute exploration as an expert knowledge acquisition method [234]

– FCA in ontology building and refining [235,79]

7.1 Attribute Exploration

Attribute exploration is an interactive knowledge acquisition procedure based on
implications and counter examples [234] that was initially applied for knowledge
acquisition in mathematics itself and still a suitable tool up to date [236].
The basic algorithm is as follows:
– Start with any (possibly empty) set of objects.
– Generate an implication valid in the current subcontext.
– If the implication is not valid in the entire context, provide an object that
violates it (a counterexample).
– Go to the next implication and so on.
A sophisticated algorithm implementation can follow the Duquenne-Guigues
base to ask minimal number of questions.
Example 29. Attribute exploration for the context of transportation means.
Whether all possible types of transportation means are enumerated with
respect to their attributes (surface, air, water, underwater)?
To this end we start attribute exploration with composing the corresponding
formal context.
underwater
surface

water
air

plane ×
amphibian car × ×
catamaran ×
car ××
submarine ××
The main steps of attribute exploration, as a dialog between system A and
expert E for transport context, is as follows:
 Introduction to Formal Concept Analysis and Its Applications in IR 85

Fig. 28. The taxonomy of transportation means as an example of not a tree-like (mul-
tiple) inheritance

– Step 1. A Question: Is it true that, when an object has attribute “Can move
underwater”, it also has attribute “Can move by water”?
– Step 1. E Answer: Yes, it is. The expert knows that it is true for submarines
and there are no other types of underwater transport.
– Step 2. A Question: Is it true that, when an object has attributes “Can move
by air” and “Can move by water” have attributes “Can move by surface”
and “Can move underwater”?
– Step 2. E Answer: No, it is not. There is a counterexample, {hydroplane}0 =
{air, water}.
– Step 3. A Question: Is it true that, when an object has attributes “Can move
by air”, “Can move by water” “Can move underwater” have attributes “Can
move by surface”?
– Step 3. E Answer: Yes, it is. {air, water, underwater}0 = ∅.
– Steps 4, 5, 6 Trivial questions.

The resulting concept lattice can be considered as a non-tree like taxonomy
of transportation means since it allows multiple inheritance in the concept hier-
archy. If the expert suppose that the not only objects but attributes are missed
then object exploration can be done in similar manner, e.g. by the same proce-
dure on the transposed context.
Exercise 23. 1. Compare the concept lattices from the previous example be-
fore starting and after completion of the attribute exploration. What is/are
new concept(s) that we have obtained? How can it/they be interpreted? 2. Per-
form attribute exploration with ConceptExplorer for a slightly modified context
from [237]
86 Dmitry I. Ignatov

Mediterranean
European
Asian
EU

G7
France ××××
Turkey × × ×
Germany ×××

7.2 FCA in ontology building and refining

Often, the notion of Ontology in Computer Science is introduced as related sets

of concepts and the typical relation can be “is-a”, “has-a”, “part-of”, or su-
per/subconcept relation. Concept lattices could be seen as ontology-like struc-
tures since they feature hierarchically related concepts by super/subconcept or-
der (cf. subsumption order in Descriptive logic). However, because of their sim-
plicity tree-like ontologies seem to be more popular, thus in the early paper of
Cimiano et al.[79], the way to transform concept lattices built from text collec-
tions to tree-like ontologies was proposed.

Exercise 24. Build concept lattice from the context of terms extracted from texts
(left). Find the transformation that resulted in the tree-like ontology of terms
on the right side.
bookable
driveable
bookable
rentable
rideable

joinable

joinable hotel rentable

hotel ×
excursion trip driveable apartment
apartment ××
car ×××
bike ×××× rideable car
excursion × ×
trip × ×
bike


Another example where FCA can help is ontology merging: The authors
of [235] successfully tested their FCA-based merging approach on two text col-
lection from touristic domain.
There is also strong connection between Description Logic, Ontologies and
Formal Concept Analysis [237].
 Introduction to Formal Concept Analysis and Its Applications in IR 87

Thus OntoComP60 [238] is a Protégé61 4 plugin for OWL ontologies com-

pletion. It enables the user to check whether an OWL ontology contains “all
relevant information” about the application domain, and extend the ontology
appropriately otherwise. It asks the users questions like “are instances of classes
C1 and C2 also instances of the class C3 ?”. If the user replies positively, then a
new axiom of the application domain (that does not follow from the ontology)
has been discovered, and this axiom should be added to the ontology. If the user
provides a counterexample to this question, i.e., an object that is an instance of
C1 , C2 and not C3 . When all such questions (about the initially given classes)
have been answered, the ontology is supposed to be complete.
Obviously, this approach that was originally introduced in [239] for complet-
ing Description Logic knowledge bases uses attribute exploration.
It seems that attribute exploration is a fruitful technique for ontology build-
ing and refinement. Two more examples, Rudolph [240] proposed its extension
for relation exploration in ontological modeling for knowledge specification and
recently in combination with machine learning techniques attribute exploration
was used for ontology refinement [241]. You probably have seen from exercise23,
that attribute exploration may be uneasy because of laborious fact checking.
However, to help potential users, in [242] the authors paired attribute explo-
ration with web information retrieval, in particular by posing appropriate queries
to search engines62 .

8 Conclusion
In the end of the invited talk at the “FCA meets IR” workshop 2013, Prof.
Carpineto has summarised strengths and limitations of FCA for IR. It seems
to be evident that IR will be increasingly relying on contextual knowledge and
structured data and FCA can improve both query pre-processing and query post-
processing of modern IR systems. Among the mentioned technologies that could
benefit from FCA are query expansion, web search diversification, ontology-
based information retrieval, querying and navigating RDF (there is a progress to
this date [243]), and many others. However, the community needs to endeavour
(by theoretical advances and system engineering) to deploy a comprehensive
FCA-based tool for information retrieval and integrate it with existing search
and indexing taking into account both the intrinsic complexity issues and the
problem of good features generation.
Even in an extensive tutorial it is not possible to cover all models and ap-
plications of Formal Concept Analysis. For example, concept lattices and its
applications in social sciences including Social Network Analysis deserve a spe-
cial treatment. The grounding steps have been done by Vincent Duquenne [85],
Linton Freeman [86] and their collaborators (see also [89] for our SNA-related
study). Another large and interesting domain is Software Engineering [244,245].
60
http://code.google.com/p/ontocomp/
61
http://www.co-ode.org/downloads/protege-x/
62
https://github.com/rjoberon/web-attribute-exploration
88 Dmitry I. Ignatov

For these two and many other topics, we also refer the readers to the recent
surveys [5,6].
Overall, we hope that this introductory material with many examples and
exercises will help the reader not only to understand the theory basics, but
having this rich variety of tools and showcases to use FCA in practice.

Acknowledgments. The author would like to thank all colleagues who have
made this tutorial possible: Jaume Baixeries, Pavel Braslavsky, Peter Becker,
Radim Belohlavek, Aliaksandr Birukou, Jean-Francois Boulicaut, Claudio Carpineto,
Florent Domenach, Fritjhof Dau, Vincent Duquenne, Bernhard Ganter, Katja
Hofmann, Robert Jaeshke, Evgenia Revne (Il’ina), Nikolay Karpov, Mehdy Kay-
toue, Sergei Kuznetsov, Rokia Missaoui, Elena Nenova, Engelbert Mephu Nguifo,
Alexei Neznanov, Lhouari Nourin, Bjoern Koester, Natalia Konstantinova, Amedeo
Napoli, Sergei Obiedkov, Jonas Poelmans, Nikita Romashkin, Paolo Rosso, Se-
bastian Rudolph, Alexander Tuzhilin, Pavel Serdyukov, Baris Serkaya, Dominik
Slezak, Marcin Szchuka, and, last but not least, the brave listeners. The author
would also like to commemorate Ilya Segalovich who inspired the author’s en-
thusiasm in Information Retrieval studies, by giving personal explanations of
near duplicate detection techniques in 2005, in particular.
Special thank should go to my grandmother, Vera, who has been hosting me
in a peaceful countryside place, Prechistoe, during the last two weeks of the final
preparations.
The author was partially supported by the Russian Foundation for Basic Re-
search grants no. 13-07-00504 and 14-01-93960 and prepared the tutorial within
the project “Data mining based on applied ontologies and lattices of closed de-
scriptions” supported by the Basic Research Program of the National Research
University Higher School of Economics.

References
1. Manning, C.D., Raghavan, P., Schütze, H.: Introduction to information retrieval.
Cambridge University Press (2008)
2. Wille, R.: Restructuring lattice theory: An approach based on hierarchies of
concepts. In Rival, I., ed.: Ordered Sets. Volume 83 of NATO Advanced Study
Institutes Series. Springer Netherlands (1982) 445–470
3. Ganter, B., Wille, R.: Formal Concept Analysis: Mathematical Foundations. 1st
edn. Springer-Verlag New York, Inc., Secaucus, NJ, USA (1999)
4. Poelmans, J., Ignatov, D.I., Viaene, S., Dedene, G., Kuznetsov, S.O.: Text mining
scientific papers: A survey on fca-based information retrieval research. In Perner,
P., ed.: ICDM. Volume 7377 of Lecture Notes in Computer Science., Springer
(2012) 273–287
5. Poelmans, J., Kuznetsov, S.O., Ignatov, D.I., Dedene, G.: Formal concept analysis
in knowledge processing: A survey on models and techniques. Expert Syst. Appl.
40(16) (2013) 6601–6623
6. Poelmans, J., Ignatov, D.I., Kuznetsov, S.O., Dedene, G.: Formal concept analysis
in knowledge processing: A survey on applications. Expert Syst. Appl. 40(16)
(2013) 6538–6560
 Introduction to Formal Concept Analysis and Its Applications in IR 89

7. Serdyukov, P., Braslavski, P., Kuznetsov, S.O., Kamps, J., Rüger, S.M., Agichtein,
E., Segalovich, I., Yilmaz, E., eds.: Advances in Information Retrieval - 35th
European Conference on IR Research, ECIR 2013, Moscow, Russia, March 24-27,
2013. Proceedings. In Serdyukov, P., Braslavski, P., Kuznetsov, S.O., Kamps, J.,
Rüger, S.M., Agichtein, E., Segalovich, I., Yilmaz, E., eds.: ECIR. Volume 7814
of Lecture Notes in Computer Science., Springer (2013)
8. Arnauld, A., Nicole, P.: Logic or the Art of Thinking, translated by Jill V.
Buroker. Cambridge University Press, (1996)
9. Birkhoff, G.: Lattice Theory (Third ed.). Providence, R.I.: Amer. Math.Soc.
(1967)
10. Ore, O.: Galois connexions. Trans. Amer. Math. Soc. 55(3) (1944) 494–513
11. Barbut, M., Monjardet, B.: Ordre et classification. Hachette, Paris (1970)
12. Duquenne, V.: Latticial structures in data analysis. Theoretical Computer Science
217(2) (1999) 407 – 436 ORDAL’96.
13. Wolski, M.: Galois connections and data analysis. Fundam. Inform. 60(1-4)
(2004) 401–415
14. Kuznetsov, S.O.: Galois connections in data analysis: Contributions from the
soviet era and modern russian research. In: Formal Concept Analysis, Foundations
and Applications. (2005) 196–225
15. Carpineto, C., Romano, G.: Concept data analysis - theory and applications.
Wiley (2005)
16. Davey, B.A., Priestley, H.A.: Introduction to Lattices and Order. 2nd edition edn.
Cambridge University Press (2002)
17. Dominich, S.: The Modern Algebra of Information Retrieval. 1 edn. Springer
Publishing Company, Incorporated (2008)
18. Wolff, K.E.: A first course in formal concept analysis. how to understand line
diagrams. In Faulbaum, F., ed.: In: Faulbaum, F. (ed.). Volume 4 of SoftStat’93.
Advances in Statistical Software. (1993) 429–438
19. Belohlávek, R.: Introduction to Formal Concept Analysis. Palacky University,
Olomouc. (2008)
20. Kuznetsov, S.O., Obiedkov, S.A.: Comparing performance of algorithms for gen-
erating concept lattices. J. Exp. Theor. Artif. Intell. 14(2-3) (2002) 189–216
21. Kourie, D.G., Obiedkov, S.A., Watson, B.W., van der Merwe, D.: An incremental
algorithm to construct a lattice of set intersections. Sci. Comput. Program. 74(3)
(2009) 128–142
22. Krajca, P., Vychodil, V.: Distributed algorithm for computing formal concepts
using map-reduce framework. In: N. Adams et al. (Eds.): IDA 2009. Volume
LNCS 5772. (2009) 333–344
23. Xu, B., de Frein, R., Robson, E., Foghlu, M.O.: Distributed formal concept
analysis algorithms based on an iterative mapreduce framework. In Domenach,
F., Ignatov, D., Poelmans, J., eds.: ICFCA 2012. Volume LNAI 7278. (2012)
292–308
24. Armstrong, W.: Dependency structures of data base relationships. Information
Processing (74) (1974) 580–583
25. Maier, D.: The Theory of Relational Databases. Computer Science Press (1983)
26. Guigues, J.L., Duquenne, V.: Familles minimales d’implications informatives rsul-
tant d’un tableau de donnes binaires. Mathmatiques et Sciences Humaines 95(1)
(1986) 5–18 In French.
27. Bazhanov, K., Obiedkov, S.A.: Optimizations in computing the duquenne-guigues
basis of implications. Ann. Math. Artif. Intell. 70(1-2) (2014) 5–24
90 Dmitry I. Ignatov

28. Baixeries, J., Kaytoue, M., Napoli, A.: Characterization of database dependencies
with FCA and pattern structures. In: Analysis of Images, Social Networks and
Texts - Third International Conference, AIST 2014, Yekaterinburg, Russia, April
10-12, 2014, Revised Selected Papers. (2014) 3–14
29. Yevtushenko, S.A.: System of data analysis ”concept explorer”. (in russian). In:
Proceedings of the 7th national conference on Artificial Intelligence KII-2000.
(2000) 127–134
30. Yevtushenko, S.: Computing and Visualizing Concept Lattices. PhD thesis, TU
Darmstadt, Fachbereich Informatik (2004)
31. Yevtushenko, S.A.: Concept Explorer. The User Guide. (September 12 2006)
32. Becker, P.: Numerical analysis in conceptual systems with toscanaj. In: Concept
Lattices, Second International Conference on Formal Concept Analysis, ICFCA
2004, Sydney, Australia, February 23-26, 2004, Proceedings. (2004) 96–103
33. Becker, P., Correia, J.H.: The toscanaj suite for implementing conceptual infor-
mation systems. In: Formal Concept Analysis, Foundations and Applications.
(2005) 324–348
34. Vogt, F., Wille, R.: TOSCANA - a graphical tool for analyzing and exploring
data. In: Graph Drawing, DIMACS International Workshop, GD ’94, Princeton,
New Jersey, USA, October 10-12, 1994, Proceedings. (1994) 226–233
35. Valtchev, P., Grosser, D., Roume, C., Hacene, M.R.: Galicia: an open platform
for lattices. In: A. de Moor B. Ganter, editor, Using Conceptual Structures:
Contributions to 11th Intl. Conference on Conceptual Structures. (2003) 241–254
36. Lahcen, B., Kwuida., L.: Lattice miner: A tool for concept lattice construction
and exploration. In: Suplementary Proceeding of International Conference on
Formal concept analysis (ICFCA’10). (2010)
37. Poelmans, J., Elzinga, P., Ignatov, D.I., Kuznetsov, S.O.: Semi-automated knowl-
edge discovery: identifying and profiling human trafficking. Int. J. General Sys-
tems 41(8) (2012) 774–804
38. Poelmans, J., Elzinga, P., Neznanov, A., Viaene, S., Kuznetsov, S., Ignatov,
D., Dedene, G.: Concept relation discovery and innovation enabling technology
(cordiet). In: Proceedings of 1st International Workshop on Concept Discovery
in Unstructured Data. Volume 757 of CEUR Workshop proceedings. (2011)
39. Neznanov, A., Ilvovsky, D., Kuznetsov, S.O.: Fcart: A new fca-based system for
data analysis and knowledge discovery. In: Contributions to the 11th International
Conference on Formal Concept Analysis, TU Dresden (2013) 31–44
40. Neznanov, A., Parinov, A.: FCA analyst session and data access tools in FCART.
In: Artificial Intelligence: Methodology, Systems, and Applications - 16th Inter-
national Conference, AIMSA 2014, Varna, Bulgaria, September 11-13, 2014. Pro-
ceedings. (2014) 214–221
41. Buzmakov, A., Neznanov, A.: Practical computing with pattern structures in
FCART environment. In: Proceedings of the International Workshop ”What can
FCA do for Artificial Intelligence?” (FCA4AI at IJCAI 2013), Beijing, China,
August 5, 2013. (2013) 49–56
42. Domenach, F.: CryptoLat – a Pedagogical Software on Lattice Cryptomorphisms
and Lattice Properties. In: Proceedings of the Tenth International Conference
on Concept Lattices and Their Applications, La Rochelle, France, October 15-18,
2013. (2013) 93–103
43. Agrawal, R., Srikant, R.: Fast algorithms for mining association rules in large
databases. In Bocca, J.B., Jarke, M., Zaniolo, C., eds.: VLDB, Morgan Kaufmann
(1994) 487–499
 Introduction to Formal Concept Analysis and Its Applications in IR 91

44. Luxenburger, M.: Implications partielles dans un contexte. Mathématiques, In-

formatique et Sciences Humaines 29(113) (1991) 35–55
45. Jäschke, R., Hotho, A., Schmitz, C., Ganter, B., Stumme, G.: Trias–an algorithm
for mining iceberg tri-lattices. In: Proceedings of the Sixth International Confer-
ence on Data Mining. ICDM ’06, Washington, DC, USA, IEEE Computer Society
(2006) 907–911
46. Ignatov, D.I., Kuznetsov, S.O., Magizov, R.A., Zhukov, L.E.: From triconcepts
to triclusters. [246] 257–264
47. Ignatov, D.I., Kuznetsov, S.O., Poelmans, J., Zhukov, L.E.: Can triconcepts be-
come triclusters? International Journal of General Systems 42(6) (2013) 572–593
48. Kuznetsov, S.O.: Machine learning and formal concept analysis. [247] 287–312
49. Ganter, B., Grigoriev, P.A., Kuznetsov, S.O., Samokhin, M.V.: Concept-based
data mining with scaled labeled graphs. In Wolff, K.E., Pfeiffer, H.D., Delugach,
H.S., eds.: ICCS. Volume 3127 of Lecture Notes in Computer Science., Springer
(2004) 94–108
50. Kuznetsov, S.O.: Fitting pattern structures to knowledge discovery in big data.
In Cellier, P., Distel, F., Ganter, B., eds.: ICFCA. Volume 7880 of Lecture Notes
in Computer Science., Springer (2013) 254–266
51. Belohlávek, R., Vychodil, V.: Discovery of optimal factors in binary data via a
novel method of matrix decomposition. J. Comput. Syst. Sci. 76(1) (2010) 3–20
52. Romashkin, N., Ignatov, D.I., Kolotova, E.: How university entrants are choosing
their department? mining of university admission process with FCA taxonomies.
In: Proceedings of the 4th International Conference on Educational Data Mining,
Eindhoven, The Netherlands, July 6-8, 2011. (2011) 229–234
53. Grigoriev, P.A., Yevtushenko, S.A.: Quda: Applying formal concept analysis in a
data mining environment. [247] 386–393
54. Han, J., Kamber., M.: Data Mining: Concepts and Techniques. Morgan Kaufmann
(2000)
55. Agrawal, R., Imieliński, T., Swami, A.: Mining association rules between sets of
items in large databases. In: ACM SIGMOD Record. Volume 22(2)., ACM (1993)
207–216
56. Pasquier, N., Bastide, Y., Taouil, R., Lakhal, L.: Efficient mining of association
rules using closed itemset lattices. Inf. Syst. 24(1) (1999) 25–46
57. Zaki, M.J., Hsiao, C.J.: Charm: An efficient algorithm for closed association rule
mining. Technical report, Computer Science, Rensselaer Polytechnic Institute
(1999)
58. Stumme, G.: Conceptual knowledge discovery with frequent concept lattices.
Technical Report FB4- Preprint 2043, TU Darmstadt (1999)
59. Stumme, G., Taouil, R., Bastide, Y., Pasquier, N., Lakhal, L.: Computing iceberg
concept lattices with T. Data Knowl. Eng. 42(2) (2002) 189–222
60. Kuznetsov, S.: Mathematical aspects of concept analysis. Journal of Mathematical
Science 80(2) (1996) 1654–1698
61. Lakhal, L., Stumme, G.: Efficient mining of association rules based on formal
concept analysis. In: Formal Concept Analysis, Foundations and Applications.
(2005) 180–195
62. Agrawal, R., Christoforaki, M., Gollapudi, S., Kannan, A., Kenthapadi, K.,
Swaminathan, A.: Mining videos from the web for electronic textbooks. In:
Formal Concept Analysis - 12th International Conference, ICFCA 2014, Cluj-
Napoca, Romania, June 10-13, 2014. Proceedings. (2014) 219–234
63. Zaki, M.J., Wagner Meira, J.: Data Mining and Analysis: Fundamental Concepts
and Algorithms. Cambridge University Press (May 2014)
92 Dmitry I. Ignatov

64. Zaki, M.J.: Spade: An efficient algorithm for mining frequent sequences. Machine
Learning 42 (2001) 31–60
65. Vander Wal, T.: Folksonomy coinage and definition. (2007)
http://vanderwal.net/folksonomy.html (accessed on 12.03.2012).
66. Mirkin, B.: Mathematical Classification and Clustering. Kluwer, Dordrecht (1996)
67. Madeira, S.C., Oliveira, A.L.: Biclustering algorithms for biological data analysis:
A survey. IEEE/ACM Trans. Comput. Biology Bioinform. 1(1) (2004) 24–45
68. Eren, K., Deveci, M., Kktun, O., atalyrek, .V.: A comparative analysis of biclus-
tering algorithms for gene expression data. Briefings in Bioinformatics (2012)
69. Besson, J., Robardet, C., Boulicaut, J.F., Rome, S.: Constraint-based concept
mining and its application to microarray data analysis. Intell. Data Anal. 9(1)
(2005) 59–82
70. Barkow, S., Bleuler, S., Prelic, A., Zimmermann, P., Zitzler, E.: Bicat: a biclus-
tering analysis toolbox. Bioinformatics 22(10) (2006) 1282–1283
71. Tarca, A.L., Carey, V.J., wen Chen, X., Romero, R., Drǎghici, S.: Machine learn-
ing and its applications to biology. PLoS Comput Biol 3(6) (June 2007) e116
72. Hanczar, B., Nadif, M.: Bagging for biclustering: Application to microarray data.
In: Machine Learning and Knowledge Discovery in Databases. Volume 6321 of
LNCS. Springer (2010) 490–505
73. Kaytoue, M., Kuznetsov, S.O., Napoli, A., Duplessis, S.: Mining gene expression
data with pattern structures in formal concept analysis. Inf. Sci. 181(10) (2011)
1989–2001
74. Blinova, V.G., Dobrynin, D.A., Finn, V.K., Kuznetsov, S.O., Pankratova, E.S.:
Toxicology analysis by means of the jsm-method. Bioinformatics 19(10) (2003)
1201–1207
75. Kuznetsov, S., Samokhin, M.: Learning closed sets of labeled graphs for chemical
applications. In: ILP 2005. Volume 3625 of LNCS (LNAI)., Springer (2005) 190–
208
76. DiMaggio, P.A., Subramani, A., Judson, R.S., Floudas, C.A.: A novel framework
for predicting in vivo toxicities from in vitro data using optimal methods for
dense and sparse matrix reordering and logistic regression. Toxicological Sciences
118(1) (2010) 251–265
77. Asses, Y., Buzmakov, A., Bourquard, T., Kuznetsov, S.O., Napoli, A.: A hybrid
classification approach based on FCA and emerging patterns – an application for
the classification of biological inhibitors. In: Proceedings of The 9th Int. Conf. on
Concept Lattices and Their Applications. (2012) 211–222
78. Dhillon, I.S.: Co-clustering documents and words using bipartite spectral graph
partitioning. In: Proceedings of the seventh ACM SIGKDD international confer-
ence on Knowledge discovery and data mining. KDD ’01, New York, NY, USA,
ACM (2001) 269–274
79. Cimiano, P., Hotho, A., Staab, S.: Learning concept hierarchies from text corpora
using formal concept analysis. J. Artif. Intell. Res. (JAIR) 24 (2005) 305–339
80. Banerjee, A., Dhillon, I.S., Ghosh, J., Merugu, S., Modha, D.S.: A Generalized
Maximum Entropy Approach to Bregman Co-clustering and Matrix Approxima-
tion. Journal of Machine Learning Research 8 (2007) 1919–1986
81. Ignatov, D.I., Kuznetsov, S.O.: Frequent itemset mining for clustering near du-
plicate web documents. [248] 185–200
82. Carpineto, C., Michini, C., Nicolussi, R.: A concept lattice-based kernel for SVM
text classification. In: ICFCA 2009. Volume LNAI 5548., Springer (2009) 237–250
 Introduction to Formal Concept Analysis and Its Applications in IR 93

83. Koester, B.: Conceptual knowledge retrieval with fooca: Improving web search
engine results with contexts and concept hierarchies. In Perner, P., ed.: Industrial
Conference on Data Mining. Volume 4065 of Lecture Notes in Computer Science.,
Springer (2006) 176–190
84. Eklund, P.W., Ducrou, J., Dau, F.: Concept similarity and related categories in
information retrieval using formal concept analysis. Int. J. General Systems 41(8)
(2012) 826–846
85. Duquenne, V.: Lattice analysis and the representation of handicap associations.
Social Networks 18(3) (1996) 217–230
86. Freeman, L.C.: Cliques, Galois lattices, and the structure of human social groups.
Social Networks 18 (1996) 173–187
87. Latapy, M., Magnien, C., Vecchio, N.D.: Basic notions for the analysis of large
two-mode networks. Social Networks 30(1) (2008) 31–48
88. Roth, C., Obiedkov, S.A., Kourie, D.G.: On Succinct Representation of Knowl-
edge Community Taxonomies with Formal Concept Analysis. Int. J. Found. Com-
put. Sci. 19(2) (2008) 383–404
89. Gnatyshak, D., Ignatov, D.I., Semenov, A., Poelmans, J.: Gaining insight in social
networks with biclustering and triclustering. In: BIR. Volume 128 of Lecture Notes
in Business Information Processing., Springer (2012) 162–171
90. du Boucher-Ryan, P., Bridge, D.G.: Collaborative Recommending using Formal
Concept Analysis. Knowl.-Based Syst. 19(5) (2006) 309–315
91. Symeonidis, P., Nanopoulos, A., Papadopoulos, A.N., Manolopoulos, Y.: Nearest-
biclusters collaborative filtering based on constant and coherent values. Inf. Retr.
11(1) (2008) 51–75
92. Ignatov, D.I., Kuznetsov, S.O.: Concept-based Recommendations for Internet
Advertisement. In Belohlavek, R., Kuznetsov, S.O., eds.: Proc. CLA 2008. Volume
Vol. 433 of CEUR WS., Palack University, Olomouc, 2008 (2008) 157–166
93. Nanopoulos, A., Rafailidis, D., Symeonidis, P., Manolopoulos, Y.: Musicbox:
Personalized music recommendation based on cubic analysis of social tags. IEEE
Transactions on Audio, Speech & Language Processing 18(2) (2010) 407–412
94. Ignatov, D.I., Nenova, E., Konstantinova, N., Konstantinov, A.V.: Boolean Matrix
Factorisation for Collaborative Filtering: An FCA-Based Approach. In: Artificial
Intelligence: Methodology, Systems, and Applications. Volume 8722 of LNCS.
Springer (2014) 47–58
95. Ignatov, D.I.: Mathematical Models, Algorithms and Software Tools of Biclus-
tering Based on Closed Sets. PhD thesis, National Research University Higher
School of Economics (2010)
96. Ignatov, D.I., Kuznetsov, S.O., Poelmans, J.: Concept-based biclustering for inter-
net advertisement. In: ICDM Workshops, IEEE Computer Society (2012) 123–130
97. Benz, D., Hotho, A., Jäschke, R., Krause, B., Mitzlaff, F., Schmitz, C., Stumme,
G.: The social bookmark and publication management system bibsonomy – A
platform for evaluating and demonstrating web 2.0 research. VLDB J. 19(6)
(2010) 849–875
98. Zhao, L., Zaki, M.J.: Tricluster: An effective algorithm for mining coherent clus-
ters in 3D microarray data. In Özcan, F., ed.: SIGMOD Conference, ACM (2005)
694–705
99. Li, A., Tuck, D.: An effective tri-clustering algorithm combining expression data
with gene regulation information. Gene regulation and systems biology 3 (2009)
49–64
100. Wille, R.: The basic theorem of Triadic Concept Analysis. Order 12 (1995)
149–158
94 Dmitry I. Ignatov

101. Lehmann, F., Wille, R.: A triadic approach to formal concept analysis. In: Pro-
ceedings of the Third International Conference on Conceptual Structures: Appli-
cations, Implementation and Theory, London, UK, Springer-Verlag (1995) 32–43
102. Krolak-Schwerdt, S., Orlik, P., Ganter, B.: Tripat: a model for analyzing three-
mode binary data. In Bock, H.H., Lenski, W., Richter, M., eds.: Information
Systems and Data Analysis. Studies in Classification, Data Analysis, and Knowl-
edge Organization. Springer Berlin Heidelberg (1994) 298–307
103. Ji, L., Tan, K.L., Tung, A.K.H.: Mining frequent closed cubes in 3d datasets. In:
Proceedings of the 32nd international conference on Very large data bases. VLDB
’06, VLDB Endowment (2006) 811–822
104. Cerf, L., Besson, J., Robardet, C., Boulicaut, J.F.: Closed patterns meet n-ary
relations. ACM Trans. Knowl. Discov. Data 3 (March 2009) 3:1–3:36
105. Cerf, L., Besson, J., Nguyen, K.N., Boulicaut, J.F.: Closed and noise-tolerant
patterns in n-ary relations. Data Min. Knowl. Discov. 26(3) (2013) 574–619
106. Georgii, E., Tsuda, K., Schölkopf, B.: Multi-way set enumeration in weight ten-
sors. Machine Learning 82(2) (2011) 123–155
107. Spyropoulou, E., De Bie, T., Boley, M.: Interesting pattern mining in multi-
relational data. Data Mining and Knowledge Discovery 28(3) (2014) 808–849
108. Voutsadakis, G.: Polyadic concept analysis. Order 19(3) (2002) 295–304
109. Ignatov, D., Gnatyshak, D., Kuznetsov, S., Mirkin, B.: Triadic formal concept
analysis and triclustering: searching for optimal patterns. Machine Learning
(2015) 1–32
110. Mirkin, B., Kramarenko, A.V.: Approximate bicluster and tricluster boxes in the
analysis of binary data. [246] 248–256
111. Gnatyshak, D., Ignatov, D.I., Kuznetsov, S.O.: From triadic fca to triclustering:
Experimental comparison of some triclustering algorithms. [249] 249–260
112. Gnatyshak, D.V., Ignatov, D.I., Kuznetsov, S.O., Nourine, L.: A one-pass triclus-
tering approach: Is there any room for big data? In: CLA 2014. (2014)
113. Ganter, B., Kuznetsov, S.O.: Hypotheses and version spaces. In de Moor, A., Lex,
W., Ganter, B., eds.: ICCS. Volume 2746 of Lecture Notes in Computer Science.,
Springer (2003) 83–95
114. Belohlávek, R., Baets, B.D., Outrata, J., Vychodil, V.: Inducing decision trees
via concept lattices. Int. J. General Systems 38(4) (2009) 455–467
115. Carpineto, C., Romano, G.: Galois: An order-theoretic approach to conceptual
clustering. In: Proceeding of ICML93, Amherst. (1993) 33–40
116. Carpineto, C., Romano, G.: A lattice conceptual clustering system and its appli-
cation to browsing retrieval. Machine Learning Vol. 24 (1996) 95–122
117. Fu, H., Fu, H., Njiwoua, P., Nguifo, E.M.: A Comparative Study of FCA-Based
Supervised Classification Algorithms. In: 2nd Int. Conf. on Formal Concept Anal-
ysis, ICFCA 2004, Sydney, Australia, February 23-26, 2004, Proceedings. (2004)
313–320
118. Rudolph, S.: Using FCA for encoding closure operators into neural networks. In:
15th International Conference on Conceptual Structures, ICCS 2007, Sheffield,
UK, July 22-27, 2007, Proceedings. (2007) 321–332
119. Tsopzé, N., Nguifo, E.M., Tindo, G.: CLANN: concept lattice-based artificial
neural network for supervised classification. In: Proceedings of the 5th Int. Conf.
on Concept Lattices and Their Applications, CLA 2007. (2007)
120. Outrata, J.: Boolean factor analysis for data preprocessing in machine learning.
In: The Ninth International Conference on Machine Learning and Applications,
ICMLA 2010, Washington, DC, USA, 12-14 December 2010. (2010) 899–902
 Introduction to Formal Concept Analysis and Its Applications in IR 95

121. Belohlávek, R., Outrata, J., Trnecka, M.: Impact of boolean factorization as
preprocessing methods for classification of boolean data. Ann. Math. Artif. Intell.
72(1-2) (2014) 3–22
122. Ganter, B., Kuznetsov, S.O.: Scale coarsening as feature selection. In: Proceed-
ings of the 6th International Conference on Formal Concept Analysis. ICFCA’08,
Berlin, Heidelberg, Springer-Verlag (2008) 217–228
123. Visani, M., Bertet, K., Ogier, J.: Navigala: an original symbol classifier based on
navigation through a Galois lattice. IJPRAI 25(4) (2011) 449–473
124. Zaki, M.J., Aggarwal, C.C.: Xrules: An effective algorithm for structural classifi-
cation of XML data. Machine Learning 62(1-2) (2006) 137–170
125. Flach, P.: Machine Learning: The Art and Science of Algorithms That Make
Sense of Data. Cambridge University Press, New York, NY, USA (2012)
126. Finn, V.: On machine-oriented formalization of plausible reasoning in f.bacon-
j.s.mill style. Semiotika i Informatika (20) (1983) 35–101 (in Russian).
127. Kuznetsov, S.: Jsm-method as a machine learning. Method. Itogi Nauki i
Tekhniki, ser. Informatika (15) (1991) 17–53 (in Russian).
128. Gusakova, S.: Paleography with jsm-method. Technical report, VINITI (2001)
129. Ganter, B., Kuznetsov, S.: Formalizing hypotheses with concepts. In Ganter,
B., Mineau, G., eds.: Conceptual Structures: Logical, Linguistic, and Computa-
tional Issues. Volume 1867 of Lecture Notes in Computer Science. Springer Berlin
Heidelberg (2000) 342–356
130. Zhuk, R., Ignatov, D.I., Konstantinova, N.: Concept learning from triadic data. In:
Proceedings of the Second International Conference on Information Technology
and Quantitative Management, ITQM 2014, National Research University Higher
School of Economics (HSE), Moscow, Russia, June 3-5, 2014. (2014) 928–938
131. Ignatov, D.I., Zhuk, R., Konstantinova, N.: Learning hypotheses from triadic la-
beled data. In: 2014 IEEE/WIC/ACM International Joint Conferences on Web
Intelligence (WI) and Intelligent Agent Technologies (IAT), Warsaw, Poland, Au-
gust 11-14, 2014 - Volume I. (2014) 474–480
132. Ganter, B., Kuznetsov, S.O.: Pattern structures and their projections. In: Concep-
tual Structures: Broadening the Base, 9th International Conference on Conceptual
Structures, ICCS 2001, Stanford, CA, USA, July 30-August 3, 2001, Proceedings.
(2001) 129–142
133. Buzmakov, A., Egho, E., Jay, N., Kuznetsov, S.O., Napoli, A., Raı̈ssi, C.: On pro-
jections of sequential pattern structures (with an application on care trajectories).
[249] 199–208
134. Kuznetsov, S.O.: Scalable knowledge discovery in complex data with pattern
structures. In: Pattern Recognition and Machine Intelligence - 5th International
Conference, PReMI 2013, Kolkata, India, December 10-14, 2013. Proceedings.
(2013) 30–39
135. Strok, F., Galitsky, B., Ilvovsky, D., Kuznetsov, S.: Pattern structure projec-
tions for learning discourse structures. In Agre, G., Hitzler, P., Krisnadhi, A.,
Kuznetsov, S., eds.: Artificial Intelligence: Methodology, Systems, and Applica-
tions. Volume 8722 of Lecture Notes in Computer Science. Springer International
Publishing (2014) 254–260
136. Belohlávek, R.: What is a fuzzy concept lattice? II. [246] 19–26
137. Kent, R.E.: Rough concept analysis: A synthesis of rough sets and formal concept
analysis. Fundam. Inform. 27(2/3) (1996) 169–181
138. Poelmans, J., Ignatov, D.I., Kuznetsov, S.O., Dedene, G.: Fuzzy and rough formal
concept analysis: a survey. Int. J. General Systems 43(2) (2014) 105–134
96 Dmitry I. Ignatov

139. Pankratieva, V.V., Kuznetsov, S.O.: Relations between proto-fuzzy concepts,

crisply generated fuzzy concepts, and interval pattern structures. Fundam. In-
form. 115(4) (2012) 265–277
140. Koren, Y., Bell, R., Volinsky, C.: Matrix factorization techniques for recommender
systems. Computer 42(8) (2009) 30–37
141. Elden, L.: Matrix Methods in Data Mining and Pattern Recognition. Society for
Industrial and Applied Mathematics (2007)
142. Hofmann, T.: Unsupervised learning by probabilistic latent semantic analysis.
Machine Learning 42(1-2) (2001) 177–196
143. Koren, Y.: Factorization meets the neighborhood: A multifaceted collaborative
filtering model. In: Proceedings of the 14th ACM SIGKDD International Confer-
ence on Knowledge Discovery and Data Mining. KDD ’08, New York, NY, USA,
ACM (2008) 426–434
144. Lin, C.J.: Projected gradient methods for nonnegative matrix factorization. Neu-
ral Comput. 19(10) (October 2007) 2756–2779
145. Nenova, E., Ignatov, D.I., Konstantinov, A.V.: An fca-based boolean matrix
factorisation for collaborative filtering. In: International Workshop FCA meets
IR at ECIR 2013. Volume Vol. 977, CEUR Workshop Proceeding. (2013) 57–73
146. Belohlávek, R., Glodeanu, C., Vychodil, V.: Optimal factorization of three-way
binary data using triadic concepts. Order 30(2) (2013) 437–454
147. Miettinen, P.: Boolean tensor factorization. In Cook, D., Pei, J., Wang, W.,
Zaı̈ane, O., Wu, X., eds.: ICDM 2011, 11th IEEE International Conference on
Data Mining, Vancouver, Canada, IEEE Computer Society, CPS (2011) 447–456
148. Ignatov, D.I., Mamedova, S., Romashkin, N., Shamshurin, I.: What can closed
sets of students and their marks say? In: Proceedings of the 4th International
Conference on Educational Data Mining, Eindhoven, The Netherlands, July 6-8,
2011. (2011) 223–228
149. Grigoriev, P., Yevtushenko, S., Grieser, G.: QuDA, a Data Miners Discovery
Environment. Technical Report AIDA-03-06, Technische Universität Darmstadt
(2003)
150. Grigoriev, P.A., Yevtushenko, S.A.: Elements of an agile discovery environment.
In: Discovery Science, 6th International Conference, DS 2003, Sapporo, Japan,
October 17-19,2003, Proceedings. (2003) 311–319
151. Grigoriev, P., Kuznetsov, S., Obiedkov, S., Yevtushenko, S.: On a version of mill’s
method of difference. In: Proc. ECAI 2002 Workshop on Concept Lattices in Data
Mining, Lyon. (2002) 26–31
152. Mooers, C.N.: A mathematical theory of language symbols in retrieval. In: Proc.
Int. Conf. Scientific Information, Washington D.C. (1958)
153. Fairthorne, R.A.: The patterns of retrieval. American Documentation 7(2) (1956)
65–70
154. Shreider, Y.: Mathematical model of classification theory. VINITI, Moscow,
(1968) 1–36 (in Russian).
155. Soergel, D.: Mathematical analysis of documentation systems. Inf. Stor. Retr.
(3) (1967) 129–173
156. Godin, R., Saunders, E., Gecsei, J.: Lattice model of browsable data spaces. Inf.
Sci. 40(2) (1986) 89–116
157. Carpineto, C., Romano, G.: Using concept lattices for text retrieval and mining.
In: Formal Concept Analysis, Foundations and Applications. (2005) 161–179
158. Priss, U.: Formal concept analysis in information science. ARIST 40(1) (2006)
521–543
 Introduction to Formal Concept Analysis and Its Applications in IR 97

159. Valverde-Albacete, F.J., Pelaez-Moreno, C.: Systems vs. methods: an analysis of

the affordances of formal concept analysis for information retrieval? In: Proc. of
Int. Workshop on FCA for IR at ECIR 2013, HSE, Moscow. (2013)
160. Ferr, S.: Camelis: Organizing and browsing a personal photo collection with a
logical information system. In Eklund, P.W., Diatta, J., Liquiere, M., eds.: CLA.
Volume 331 of CEUR Workshop Proceedings., CEUR-WS.org (2007)
161. Hotho, A., Jäschke, R., Schmitz, C., Stumme, G.: Information retrieval in folk-
sonomies: Search and ranking. In Sure, Y., Domingue, J., eds.: ESWC. Volume
4011 of Lecture Notes in Computer Science., Springer (2006) 411–426
162. Ignatov, D.I., Konstantinov, A.V., Chubis, Y.: Near-duplicate detection for
online-shops owners: An fca-based approach. [7] 722–725
163. Kuznetsov, S.O., Ignatov, D.I.: Concept stability for constructing taxonomies of
web-site users. in Proc. Social Network Analysis and Conceptual Structures: Ex-
ploring Opportunities, S. Obiedkov, C. Roth (Eds.), Clermont-Ferrand (France),
February 16, 2007 (2007)
164. Ignatov, D.I., Kaminskaya, A.Y., Bezzubtseva, A.A., Konstantinov, A.V., Poel-
mans, J.: Fca-based models and a prototype data analysis system for crowdsourc-
ing platforms. [250] 173–192
165. Carpineto, C., Romano, G.: A survey of automatic query expansion in information
retrieval. ACM Comput. Surv. 44(1) (2012) 1
166. Carpineto, C., Romano, G.: Order-theoretical ranking. JASIS 51(7) (2000) 587–
601
167. Carpineto, C., Osinski, S., Romano, G., Weiss, D.: A survey of web clustering
engines. ACM Comput. Surv. 41(3) (2009)
168. Carpineto, C., Romano, G.: Exploiting the potential of concept lattices for infor-
mation retrieval with CREDO. J. UCS 10(8) (2004) 985–1013
169. Ducrou, J., Eklund, P.W.: Searchsleuth: The conceptual neighbourhood of an
web query. In: Proceedings of the Fifth International Conference on Concept
Lattices and Their Applications, CLA 2007, Montpellier, France, October 24-26,
2007. (2007)
170. Dau, F., Ducrou, J., Eklund, P.W.: Concept similarity and related categories in
searchsleuth. In: Conceptual Structures: Knowledge Visualization and Reasoning,
16th International Conference on Conceptual Structures, ICCS 2008, Toulouse,
France, July 7-11, 2008, Proceedings. (2008) 255–268
171. Nauer, E., Toussaint, Y.: Crechaindo: an iterative and interactive web information
retrieval system based on lattices. Int. J. General Systems 38(4) (2009) 363–378
172. Kim, M., Compton, P.: Evolutionary document management and retrieval for
specialized domains on the web. Int. J. Hum.-Comput. Stud. 60(2) (2004) 201–
241
173. Kim, M., Compton, P.: A hybrid browsing mechanism using conceptual scales.
In: Advances in Knowledge Acquisition and Management, Pacific Rim Knowledge
Acquisition Workshop, PKAW 2006, Guilin, China, August 7-8, 2006, Revised
Selected Papers. (2006) 132–143
174. Cigarrán, J.M., Gonzalo, J., Peñas, A., Verdejo, F.: Browsing search results via
formal concept analysis: Automatic selection of attributes. [247] 74–87
175. Cole, R.J., Eklund, P.W., Stumme, G.: Document retrieval for e-mail search
and discovery using formal concept analysis. Applied Artificial Intelligence 17(3)
(2003) 257–280
176. Cole, R.J., Eklund, P.W.: Browsing semi-structured web texts using formal con-
cept analysis. In: Conceptual Structures: Broadening the Base, 9th International
98 Dmitry I. Ignatov

Conference on Conceptual Structures, ICCS 2001, Stanford, CA, USA, July 30-
August 3, 2001, Proceedings. (2001) 319–332
177. Eklund, P.W., Cole, R.J.: A knowledge representation for information filtering
using formal concept analysis. Electron. Trans. Artif. Intell. 4(C) (2000) 51–61
178. Eklund, P.W., Ducrou, J., Brawn, P.: Concept lattices for information visualiza-
tion: Can novices read line-diagrams? [247] 57–73
179. Eklund, P.W., Wormuth, B.: Restructuring help systems using formal concept
analysis. In: Formal Concept Analysis, Third International Conference, ICFCA
2005, Lens, France, February 14-18, 2005, Proceedings. (2005) 129–144
180. Stojanovic, N.: On the query refinement in the ontology-based searching for
information. Inf. Syst. 30(7) (2005) 543–563
181. Spyratos, N., Meghini, C.: Preference-based query tuning through refine-
ment/enlargement in a formal context. In: Foundations of Information and Knowl-
edge Systems, 4th International Symposium, FoIKS 2006, Budapest, Hungary,
February 14-17, 2006, Proceedings. (2006) 278–293
182. Grand, B.L., Aufaure, M., Soto, M.: Semantic and conceptual context-aware
information retrieval. In: Advanced Internet Based Systems and Applications,
Second International Conference on Signal-Image Technology and Internet-Based
Systems, SITIS 2006, Hammamet, Tunisia, December 17-21, 2006, Revised Se-
lected Papers. (2006) 247–258
183. Eklund, P.W., Ducrou, J.: Navigation and annotation with formal concept analy-
sis. In: Knowledge Acquisition: Approaches, Algorithms and Applications, Pacific
Rim Knowledge Acquisition Workshop, PKAW 2008, Hanoi, Vietnam, December
15-16, 2008, Revised Selected Papers. (2008) 118–121
184. Cigarrán, J.M., Peñas, A., Gonzalo, J., Verdejo, F.: Automatic selection of noun
phrases as document descriptors in an fca-based information retrieval system. In:
Formal Concept Analysis, Third International Conference, ICFCA 2005, Lens,
France, February 14-18, 2005, Proceedings. (2005) 49–63
185. Recio-Garcı́a, J.A., Gómez-Martı́n, M.A., Dı́az-Agudo, B., González-Calero, P.A.:
Improving annotation in the semantic web and case authoring in textual CBR.
In: Advances in Case-Based Reasoning, 8th European Conference, ECCBR 2006,
Fethiye, Turkey, September 4-7, 2006, Proceedings. (2006) 226–240
186. Liu, M., Shao, M., Zhang, W., Wu, C.: Reduction method for concept lattices
based on rough set theory and its application. Computers & Mathematics with
Applications 53(9) (2007) 1390–1410
187. Lungley, D., Kruschwitz, U.: Automatically maintained domain knowledge: Initial
findings. In: Advances in Information Retrieval, 31th European Conference on
IR Research, ECIR 2009, Toulouse, France, April 6-9, 2009. Proceedings. (2009)
739–743
188. Ahmad, I., Jang, T.: Old fashion text-based image retrieval using FCA. In: ICIP
(3). (2003) 33–36
189. Ducrou, J., Vormbrock, B., Eklund, P.W.: Fca-based browsing and searching of
a collection of images. In: Conceptual Structures: Inspiration and Application,
14th International Conference on Conceptual Structures, ICCS 2006, Aalborg,
Denmark, July 16-21, 2006, Proceedings. (2006) 203–214
190. Ducrou, J.: Dvdsleuth: A case study in applied formal concept analysis for nav-
igating web catalogs. In: Conceptual Structures: Knowledge Architectures for
Smart Applications, 15th International Conference on Conceptual Structures,
ICCS 2007, Sheffield, UK, July 22-27, 2007, Proceedings. (2007) 496–500
 Introduction to Formal Concept Analysis and Its Applications in IR 99

191. Amato, G., Meghini, C.: Faceted content-based image retrieval. In: 19th Interna-
tional Workshop on Database and Expert Systems Applications (DEXA 2008),
1-5 September 2008, Turin, Italy. (2008) 402–406
192. Ferré, S.: Camelis: a logical information system to organise and browse a collection
of documents. Int. J. General Systems 38(4) (2009) 379–403
193. Poelmans, J., Elzinga, P., Viaene, S., Dedene, G.: Formally analysing the concepts
of domestic violence. Expert Syst. Appl. 38(4) (2011) 3116–3130
194. Wolff, K.E.: States, transitions, and life tracks in temporal concept analysis. In:
Formal Concept Analysis, Foundations and Applications. (2005) 127–148
195. Elzinga, P., Poelmans, J., Viaene, S., Dedene, G., Morsing, S.: Terrorist threat
assessment with formal concept analysis. In: IEEE International Conference on
Intelligence and Security Informatics, ISI 2010, Vancouver, BC, Canada, May
23-26, 2010, Proceedings. (2010) 77–82
196. Elzinga, P., Wolff, K.E., Poelmans, J.: Analyzing chat conversations of pedophiles
with temporal relational semantic systems. In: 2012 European Intelligence and
Security Informatics Conference, EISIC 2012, Odense, Denmark, August 22-24,
2012. (2012) 242–249
197. Bullens, R., Van Horn, J.: Daad uit liefde: Gedwongen prostitutie van jonge
meisjes. Justitiele Verkenningen 26(6) (2000) 25–41
198. Koester, B., Schmidt, S.: Information superiority via formal concept analysis. In
Argamon, S., Howard, N., eds.: Computational Methods for Counterterrorism.
Springer Berlin Heidelberg (2009) 143–171
199. Obiedkov, S.A., Kourie, D.G., Eloff, J.H.P.: Building access control models with
attribute exploration. Computers & Security 28(1-2) (2009) 2–7
200. Dau, F., Knechtel, M.: Access policy design supported by FCA methods. [248]
141–154
201. Zhukov, L.E.: Spectral clustering of large advertiser datasets. Technical report,
Overture R&D (April 2004)
202. Sarwar, B.M., Karypis, G., Konstan, J.A., Riedl, J.: Analysis of recommendation
algorithms for e-commerce. In: ACM Conference on Electronic Commerce. (2000)
158–167
203. Besson, J., Robardet, C., Boulicaut, J.F., Rome, S.: Constraint-based bi-set min-
ing for biologically relevant pattern discovery in microarray data. Intelligent Data
Analysis journal 9(1) (2005) 59–82
204. Szathmary, L., Napoli, A.: CORON: A Framework for Levelwise Itemset Mining
Algorithms. In: Suppl. Proc. of ICFCA ’05, Lens, France. (Feb 2005) 110–113
205. Szathmary, L., Napoli, A., Kuznetsov, S.O.: ZART: A Multifunctional Itemset
Mining Algorithm. In: Proc. of the 5th Intl. Conf. on Concept Lattices and Their
Applications (CLA ’07), Montpellier, France (Oct 2007) 26–37
206. Crystal, D.: A dictionary of linguistics and phonetics. third edn. Oxford: Blackwell
Publishers (1991)
207. Symeonidis, P., Ruxanda, M.M., Nanopoulos, A., Manolopoulos, Y.: Ternary
semantic analysis of social tags for personalized music recommendation. In Bello,
J.P., Chew, E., Turnbull, D., eds.: ISMIR. (2008) 219–224
208. Alqadah, F., Reddy, C., Hu, J., Alqadah, H.: Biclustering neighborhood-based
collaborative filtering method for top-n recommender systems. Knowledge and
Information Systems (2014) 1–17
209. Adomavicius, G., Tuzhilin, A.: Toward the next generation of recommender sys-
tems: A survey of the state-of-the-art and possible extensions. IEEE Trans. on
Knowl. and Data Eng. 17(6) (June 2005) 734–749
100 Dmitry I. Ignatov

210. Ignatov, D.I., Poelmans, J., Dedene, G., Viaene, S.: A new cross-validation tech-
nique to evaluate quality of recommender systems. In Kundu, M., Mitra, S.,
Mazumdar, D., Pal, S., eds.: Perception and Machine Intelligence. Volume 7143
of Lecture Notes in Computer Science. Springer Berlin Heidelberg (2012) 195–202
211. Brin, S., Davis, J., Garcı́a-Molina, H.: Copy detection mechanisms for digital
documents. SIGMOD Rec. 24(2) (May 1995) 398–409
212. Broder, A.Z., Glassman, S.C., Manasse, M.S., Zweig, G.: Syntactic clustering of
the web. Computer Networks 29(8-13) (1997) 1157–1166
213. Ilyinsky, S., Kuzmin, M., Melkov, A., Segalovich, I.: An efficient method to detect
duplicates of web documents with the use of inverted index. In: Proc. 11th Int.
World Wide Web Conference (WWW’2002), Honolulu, Hawaii, USA, 7-11 May
2002, ACM (2002)
214. Broder, A.Z., Charikar, M., Frieze, A.M., Mitzenmacher, M.: Min-wise indepen-
dent permutations (extended abstract). In: Proceedings of the Thirtieth Annual
ACM Symposium on the Theory of Computing, Dallas, Texas, USA, May 23-26,
1998. (1998) 327–336
215. Broder, A.Z.: Identifying and filtering near-duplicate documents. In: Combinato-
rial Pattern Matching, 11th Annual Symposium, CPM 2000, Montreal, Canada,
June 21-23, 2000, Proceedings. (2000) 1–10
216. Grahne, G., Zhu, J.: Efficiently using prefix-trees in mining frequent itemsets. In:
FIMI ’03, Frequent Itemset Mining Implementations, Proceedings of the ICDM
2003 Workshop on Frequent Itemset Mining Implementations, 19 December 2003,
Melbourne, Florida, USA. (2003)
217. Karypis, G.: Cluto. a clustering toolkit. Technical Report: 2-017 MN 55455, Uni-
versity of Minnesota, Department of Computer Science Minneapolis (November
28 2003)
218. Potthast, M., Stein, B.: New issues in near-duplicate detection. In: Data Analysis,
Machine Learning and Applications - Proceedings of the 31st Annual Conference
of the Gesellschaft für Klassifikation e.V., Albert-Ludwigs-Universität Freiburg,
March 7-9, 2007. (2007) 601–609
219. Zelenkov, Y.G., Segalovich, I.V.: Comparative analysis of near-duplicate detection
methods of web documents. In: Proc. 9th All-Russian Scientific Conference Digital
Libraries: Advanced Methods and Technologies, Digital Collections, Pereslavl-
Zalessky. (2007) 166–174 (in Russian).
220. Ignatov, D.I., Jánosi-Rancz, K.T., Kuznetzov, S.O.: Towards a framework for
near-duplicate detection in document collections based on closed sets of attributes.
Acta Universitatis Sapientiae. Informatica 1(2) (2009) 215–233
221. Ignatov, D., Kuznetsov, S., Lopatnikova, V., Selitskiy, I.: Development and apro-
bation of near duplicate detection system for collections of r&d documents. Busi-
ness informatics (4) (2008) 21–28 (in Russian).
222. Ley, M.: DBLP - some lessons learned. PVLDB 2(2) (2009) 1493–1500
223. Benz, D., Hotho, A., Jäschke, R., Krause, B., Stumme, G.: Query logs as folk-
sonomies. Datenbank-Spektrum 10(1) (2010) 15–24
224. Doerfel, S., Jäschke, R.: An analysis of tag-recommender evaluation procedures.
In: Seventh ACM Conference on Recommender Systems, RecSys ’13, Hong Kong,
China, October 12-16, 2013. (2013) 343–346
225. Kuznetsov, S.O., Ignatov, D.I.: Concept stability for constructing taxonomies of
web-site users. In: Proc. Social Network Analysis and Conceptual Structures: Ex-
ploring Opportunities, S. Obiedkov, C. Roth (Eds.), Clermont-Ferrand (France),
February 16, 2007. (2007) 19–24
 Introduction to Formal Concept Analysis and Its Applications in IR 101

226. Kuznetsov, S.: Stability as an estimate of the degree of substantiation of hy-

potheses derived on the basis of operational similarity. Nauchn. Tekh. Inf., Ser.2
(Automat. Document. Math. Linguist.) (12) (1990) 21–29
227. Kuznetsov, S.O.: On stability of a formal concept. Ann. Math. Artif. Intell.
49(1-4) (2007) 101–115
228. Roth, C., Cointet, J.P.: Social and semantic coevolution in knowledge networks.
Social Networks 32 (2010) 16–29
229. Yavorsky, R.: Research Challenges of Dynamic Socio-Semantic Networks. In
Ignatov, D., Poelmans, J., Kuznetsov, S., eds.: CEUR Workshop proceedings Vol-
757, CDUD’11 - Concept Discovery in Unstructured Data. (2011) 119–122
230. Howe, J.: The rise of crowdsourcing. Wired (2006)
231. Ignatov, D.I., Mikhailova, M., (Kaminskaya), A.Y.Z., Malioukov, A.: Recommen-
dation of ideas and antagonists for crowdsourcing platform witology. In: Proc. of
8th RuSSIR, Springer (2014) (this volume).
232. Ignatov, D.I., Kaminskaya, A.Y., Konstantinova, N., Malyukov, A., Poelmans,
J.: Fca-based recommender models and data analysis for crowdsourcing platform
witology. In: Graph-Based Representation and Reasoning - 21st International
Conference on Conceptual Structures, ICCS 2014, Iaşi, Romania, July 27-30,
2014, Proceedings. (2014) 287–292
233. Ignatov, D.I., Kaminskaya, A.Y., Konstantinova, N., Konstantinov, A.V.: Recom-
mender system for crowdsourcing platform witology. In: 2014 IEEE/WIC/ACM
International Joint Conferences on Web Intelligence (WI) and Intelligent Agent
Technologies (IAT), Warsaw, Poland, August 11-14, 2014 - Volume II. (2014)
327–335
234. Ganter, B.: Attribute exploration with background knowledge. Theoretical Com-
puter Science 217(2) (1999) 215 – 233 ORDAL’96.
235. Stumme, G., Maedche, A.: Fca-merge: Bottom-up merging of ontologies. In
Nebel, B., ed.: IJCAI, Morgan Kaufmann (2001) 225–234
236. Revenko, A., Kuznetsov, S.O.: Attribute exploration of properties of functions
on sets. Fundam. Inform. 115(4) (2012) 377–394
237. Sertkaya, B.: A survey on how description logic ontologies benefit from FCA. In:
Proceedings of the 7th International Conference on Concept Lattices and Their
Applications, Sevilla, Spain, October 19-21, 2010. (2010) 2–21
238. Sertkaya, B.: Ontocomp: A protégé plugin for completing OWL ontologies. In:
The Semantic Web: Research and Applications, 6th European Semantic Web Con-
ference, ESWC 2009, Heraklion, Crete, Greece, May 31-June 4, 2009, Proceedings.
(2009) 898–902
239. Baader, F., Ganter, B., Sertkaya, B., Sattler, U.: Completing description logic
knowledge bases using formal concept analysis. In: IJCAI 2007, Proceedings of the
20th International Joint Conference on Artificial Intelligence, Hyderabad, India,
January 6-12, 2007. (2007) 230–235
240. Rudolph, S.: Relational exploration: combining description logics and formal
concept analysis for knowledge specification. PhD thesis, Dresden University of
Technology (2006)
241. Potoniec, J., Rudolph, S., Lawrynowicz, A.: Towards combining machine learning
with attribute exploration for ontology refinement. In: Proceedings of the ISWC
2014 Posters & Demonstrations Track a track within the 13th International Se-
mantic Web Conference, ISWC 2014, Riva del Garda, Italy, October 21, 2014.
(2014) 229–232
102 Dmitry I. Ignatov

242. Jäschke, R., Rudolph, S.: Attribute exploration on the web. In Cellier, P., Distel,
F., Ganter, B., eds.: Contributions to the 11th International Conference on Formal
Concept Analysis, Technische Universitt Dresden (May 2013) 19–34
243. Codocedo, V., Lykourentzou, I., Napoli, A.: A semantic approach to concept
lattice-based information retrieval. Ann. Math. Artif. Intell. 72(1-2) (2014) 169–
195
244. Tilley, T., Cole, R., Becker, P., Eklund, P.W.: A survey of formal concept anal-
ysis support for software engineering activities. In: Formal Concept Analysis,
Foundations and Applications. (2005) 250–271
245. Arévalo, G., Desnos, N., Huchard, M., Urtado, C., Vauttier, S.: Formal concept
analysis-based service classification to dynamically build efficient software com-
ponent directories. Int. J. General Systems 38(4) (2009) 427–453
246. Kuznetsov, S.O., Slezak, D., Hepting, D.H., Mirkin, B., eds.: Rough Sets, Fuzzy
Sets, Data Mining and Granular Computing - 13th International Conference,
RSFDGrC 2011, Moscow, Russia, June 25-27, 2011. Proceedings. In Kuznetsov,
S.O., Slezak, D., Hepting, D.H., Mirkin, B., eds.: RSFDGrC. Volume 6743 of
Lecture Notes in Computer Science., Springer (2011)
247. Eklund, P.W., ed.: Concept Lattices, Second International Conference on For-
mal Concept Analysis, ICFCA 2004, Sydney, Australia, February 23-26, 2004,
Proceedings. In Eklund, P.W., ed.: ICFCA. Volume 2961 of Lecture Notes in
Computer Science., Springer (2004)
248. Rudolph, S., Dau, F., Kuznetsov, S.O., eds.: Conceptual Structures: Leveraging
Semantic Technologies, 17th International Conference on Conceptual Structures,
ICCS 2009, Moscow, Russia, July 26-31, 2009. Proceedings. In Rudolph, S., Dau,
F., Kuznetsov, S.O., eds.: ICCS. Volume 5662 of Lecture Notes in Computer
Science., Springer (2009)
249. Ojeda-Aciego, M., Outrata, J., eds.: Proceedings of the Tenth International Con-
ference on Concept Lattices and Their Applications, La Rochelle, France, October
15-18, 2013. In Ojeda-Aciego, M., Outrata, J., eds.: CLA. Volume 1062 of CEUR
Workshop Proceedings., CEUR-WS.org (2013)
250. Pfeiffer, H.D., Ignatov, D.I., Poelmans, J., Gadiraju, N., eds.: Conceptual Struc-
tures for STEM Research and Education, 20th International Conference on Con-
ceptual Structures, ICCS 2013, Mumbai, India, January 10-12, 2013. Proceedings.
Volume 7735 of Lecture Notes in Computer Science., Springer (2013)