Recommender Systems and Their Fairness for User Preferences: A

Literature Study
Osman Ali Sadek Ibrahima,1,∗, Eman M. G. Younisb
a
Computer Science Department, Faculty of Science, Minia University, Egypt
b
Faculty of Computers and Information, Minia University, Egypt

Abstract
Recommender System (RS) is an information system that provides suggestions for items of
information to be used by their users. These suggestions can improve the user choices for
typically needed items. RS proved its usefulness in commercial environments such as Amazon
and also proved its importance in scientific environments such as ScienceDirect, Citeseer,
among others. Nowadays, RS is used extensively in social media environments such as
Facebook, Twitter and LinkedIn. However, the methods used for ranking the recommended
items of information and data may be biased and directing users towards unfair decisions.
In this literature study, we introduce a background knowledge about most of RS techniques
mentioned in the literature. Then, we identify the limitations of each technique that may
be the reason for introducing biased recommendations to the user.
Keywords: Recommender Systems, Fairness, Information Systems, Bias

∗
I am corresponding author
Email addresses: [email protected] (Osman Ali Sadek Ibrahim), [email protected]
(Eman M. G. Younis)
Preprint submitted to Information Systems Journal in 1/11/2018 November 17, 2018
1. Introduction
Before purchasing things or making decisions, we usually ask our friends or relatives
about their suggestions and recommendations. Nowadays, most of our deals and decisions
have been accomplished using digital environments such as computers, laptops and mobile
phones over the web. However, the rapid increase in the size of data and its availability on
the web boosted the need for online decision support and recommendation systems. Further-
more, social networks such as Twitter, Facebook and LinkedIn can be considered as efficient
tools to direct the user populations to specific objectives according to their opinion trends.
The main objective of these Recommender Systems (RSs) is to provide relevant suggestions
to online users. These suggestions help users to make their decisions as the the best choices
from available alternative items. In other words, RSs are mainly directing their users, who
have lack sufficient experiences, towards the better choices among many alternatives. The
success and popularity of these RSs in social networks and economic environments inspired
using them in other domains such as in Healthcare, Banking and Travel recommendations.

The previous discussion mentioned the general definition for RS, its importance and ob-
jectives. Figure 1 shows the main RS types that in the literature study. From this figure, we
can notice that there are 3 main categories for the RSs which are: 1) Collaborative Filtering,
2) Personalised Recommender System and 3) Content-Based Recommender System. Each
of these categories contains sub-types and these types are discussed in some details in the
following sections.

This literature study is organised as follows. Section 2 discusses the Collaborative

Filtering (CF) methods. Section 3 presents Personalised techniques for RSs. The Content-
Based RSs is discussed in section 4, while section 5 states the robustness techniques for
RSs. Section 6 provides the conclusions and summary.

2
3
Figure 1: Main Recommender System Types
2. Collaborative Filtering Recommender Systems (CF-RSs)
CF-RSs are the most well-known RSs on the web. In this type of RSs, filtering items
from a large set of alternatives is done by similar user preferences collaboratively. The main
concept of CF-RSs is that if two users have the same interests in the past, they may also
have the same interests in the future. On the other hand, if user A and user B purchased
similar items in the past, while user A recently purchase IPhone mobile, which user B has
not seen nor purchased it yet, then the idea is to recommend this unseen new item to user B.
The CF methods are based on the inter-item correlation or inter-user correlation methods.
Furthermore, CF methods can be divided into two categories which are as follows:

• Neighbourhood-Based technique which is also known as Memory-Based technique. In

this technique, the recommendation is based on the user or item of information neigh-
bourhoods. These neighbourhoods are determined based on the similarity between the
active user or the selected item with other users or unseen item of information.

• Model-Based technique which is using data mining and machine learning techniques.
They are also called predictive models. Some examples of such model-based methods
include Random Forest (RF), Bayesian model, Latent Semantic Analysis and Rule-
Based Models.

On the other hand, one of the main differences between the model-based technique and
the memory-based technique is the use of the training phase to produce the predictive rec-
ommended items. The training phase in the model-based technique consumes a significant
amount of time to produce the predictive model. This model is used to predict and recom-
mend the item of information to RSs users. On the other hand, the memory-based technique
does not have a training phase. It is only predicting and recommending the items of infor-
mation. Thus, the model-based technique is the most suitable technique for real-time with
limited available resources. In the following subsection, we will illustrate both techniques in
some details.

2.1. Neighbourhood-Based (Memory-Based) Recommender Systems

Neighbourhood-based techniques for RSs are also called Memory-based techniques. In
this type, the similarity values between RSs items or users are used to recommend the unseen
items of information or product to the users based on the neighbourhood similarities.

2.1.1. Nearest Neighbourhood

User-based CF
In this technique, the recommended item is introduced to an active user A based on
other user votes for the items [1]. Assuming that vk,i is the corresponding vote for user k

4
for item i and Ik is the set of items that user k has voted for them. The mean vote for user
k can be represented by:
1 X
v¯k = vk,i (1)
Ik i∈I
k

The predictive vote value to recommend unseen item i to an active user A (PA,i ) is the
weighted-sum value of other user votes. This value can be calculated by:
n
X
PA,i = v¯k + C w(A, k)(vk,i − v¯k ) (2)
k=1

where C isP a normalisation factor that is used to normalise the total weight sum to
unity (i.e. C nk=1 w(A, k) = 1). On the other hand, w(A, k) is the similarity or the
correlation weight value between an active user A and user k. There are several similarities
and correlation metrics can be used as the relationship between the active user and other
users. Examples of these metrics are Cosine and Pearson similarity among others. The
details of similarity and correlation functions are presented in [1, 2]. Furthermore, additional
extensions for calculating w(A, k) are Inverse User Frequency and Case Amplification as were
mentioned in [1].

Item-based CF
In this technique, the RSs determine the recommended items to the users based on the
similarity or correlation of the selected (active) items to the available unseen items [3, 4].
These similarities or correlations are measured based on the collaborative user rates of the
previously selected items to the new unseen available items. Assuming that the user rating
→
−
vector for selected item i by active user A is i and the user rating vector for candidate
→
−
item j that may be recommended to active user A is j . Thus, the probability of Cosine
similarity or Pearson correlation chance simi,j or corri,j that item j can be recommended
to user A can be calculated by:
→
− → −
→
− → − i · j
simi,j = cosine( i , j ) = →− →
− (3)
|| i ||X|| j ||
OR
P
− r¯i )(ru,j − r¯j )
u∈Ui,j (ru,i
corri,j = qP P (4)
− 2. 2
u∈Ui (ru,i r
¯i ) u∈Uj (ru,j − r¯i )

where Ui,j is the set of users who voted for both items i and j, while Ui and Uj are the
users voted for item i and item j consequently. Furthermore, ru,i and ru,j are the vote or
→
− →
−
user rates for items i and j which are represented by vote or rate vector i and j . On the
other hand, the mean rate value of an item is represented by r̄.

5
2.1.2. Social Network Traversal (SNT)
In this techniques the users in the social network were asked to provide a recommendation
and rating for items of information or products to the active user [5]. The SNT technique
can be divided into three categories. The first category is Trust Weighted Prediction method
[2] and the second category is Bayesian Inference Based Prediction method [6], while the
third category is Random Walk Based Approach method [7]. The SNT technique is based
on the trust-aware approaches that will be discussed in section 5.2.

2.2. Model-Based Recommender Systems

2.2.1. Machine Learning Techniques
Clustering-based CF
Clustering is a technique to gather a set of similar items together based on their similar
CF relevance feedback between users. However, clustering technique based on CF data
method is mainly about clustering users based on their selection activity of the users on the
RSs. The users in this technique are vector of rating features and each feature represents
an item of information or a product or a movie based on the RS type. Then, similarity
matching measures as mentioned in 2.1.1 are used to gather the similar users into groups.
Finally, the groups or the clusters are used to recommend items of information or products
for a user based on other users existing in her/his cluster. In this category, K-means is the
most well-known clustering method used in this technique.

2.2.2. Matrix Factorisation (MF) Technique

Matrix Factorisation (MF) was the first technique used to develop CF-RS. This technique
is much similar to Vector Space Model (VSM). This technique represents each user by a
vector of rates for the RS’s items and each RS’s item by a vector of user rates. Then,
the technique uses a users-items matrix decomposition to the lower dimension size. This
technique is much similar to the linear regression method by proposing a linear equation
to represent the relationship between the mean user rates with the user bias and item bias
values to produce the predicted rates for the RS’s items. The MF technique produced divided
into two mean methods which are: 1) Probabilistic MF method [8, 9] and 2) Similarity-
Based method [10, 11, 12]. In the first method, the MF equation proposed based on the
probabilistic distribution of the user rates, user bias and item rate bias, while the second
method is based on the similarity between users and RS’s items with considering a Singular
Value Decomposition for the user-item matrix of rates. The most well-known MF technique
is Singular Value Decomposition method which is implemented in Surprise Python Package
[13].

3. Personalised Recommender Systems (PRS)

In PRS, the recommender systems build their recommendation models based on indi-
vidual user relevance feedback. The RS starts with recommendation models created by CF
relevance feedback to remove the barrier of user cold start. This is because, at the begin-
ning of user interaction, there are no enough relevance feedback labels provided by the users.
6
Thus, the machine learning technique mentioned in Section 4.2.2 can learn at the beginning
from CF labels until having enough personalised relevance labels. Recently, there are several
recommender systems and search engines starting to use machine learning techniques, but
there is no clear idea if they will use personalised and CF labels to customise the recom-
mended item based only on self-user taste or not. Thus, there is no relevant literature about
this new technique.

3.1. Context-aware Personalised Recommender Systems

In this technique, RSs are adapted systems in which the recommended items are sent to
the users based on the context of the user environment. For example, user A may like ice
cream, but he has now a cold. Thus, it should not recommend an ice cream for the user
at the moment. Various other examples of the context may exist in which the RSs change
their type of recommendation based on the user requirements such as pregnancy or selling
trends among other of environments. More details about Context-aware RSs in [14].

4. Content-Based Recommender Systems (CB-RSs)

The main difference between CB-RSs and CF-RSs is that CB-RSs rely on the features
of items and users for predictions, while CF-RSs only uses the user rating data for items
to make predictions. The CB-RSs has the root of research that belongs to Information
Retrieval (IR) research field. This is because CB-RSs and IR are mainly about an analysis
of the data content to recommend information of items to the users.

4.1. Mathematical Models (MM)

A MM model refers to the way in which the recommender system organises, indexes
and recommends information. There are two prominent MM model categories: the first is
known as Vector Space Model (VSM), the second is called Probabilistic model [15]. All
these models use the same procedure for web-pages pre-processing and they differ in assign-
ing weights for each word and similarity measures between the selected items and unseen
items for recommending similar items.

4.1.1. Vector Space Model (VSM)

Vector Space Model (VSM) is a type of TVM representation. In VSM, the web-pages
are represented as vectors in the web-pages space [15, 16]. Each dimension in the web-pages
space represents the weights given to the term in the web-pages collection. In other words,
the web-page in the IR collection contains text words. After the pre-processing procedure,
we obtain index terms or keywords representing each web-page. Then, we assign weights for
each index term in the web-page collection. This weight represents the importance or the
information content of that index term in a given web-page. Each web-page is stored in the
following form:

d = (w1 , w2 , ..., wn )
7
 where d is a web-page in the web-page collection, wi is the weight value of term i in the
word collection and n is the number of words in the word collection that represent the infor-
mation content of the web-page collection. The word weight can be assigned statistically or
manually by trained indexer with expertise in the content of the web-page collections. When
users selected some web-pages as textual data relevant to them, Then the RSs measure the
similarity between the selected web-pages and the unseen web-pages to recommend some of
them based on the similarity values.

Term-Weighting Schemes in VSM

A ”good” index term is a word that has a high discrimination value or weight that
decreases the similarity between dissimilar web-pages when assigned to the collection [16].
The simple word or term weighting scheme uses the number of term occurrences in a given
web-page which is called term frequency (tf). However, there is a drawback in this scheme.
It may be that the term gets high weight value in every web-page at the same time because
the term is repeated in every web-page and this makes it not a good discriminator term
for web-pages. Sparck Jones [17] proposed another weighting scheme called Inverse Docu-
ment Frequency (idf) represented as log (N/n) where N is the total number of web-pages
or documents in the collection and n is the number of documents to which a term is assigned.

Salton and Buckley [18] proposed several weighting schemes for automatic text retrieval.
Salton and Buckley classified a term weighting scheme according to three main components:
term frequency, collection frequency and normalisation components. One of these combi-
nations is Term Frequency-Inverse Document Frequency (TF-IDF). The TF-IDF weighting
scheme is now the most well-known term-weighting scheme in VSM that has been widely
used in the literature such as in [19, 20, 21].

Similarity Matching Between Selected Web-pages and Unseen Web-pages in VSM

Once the web-pages vectors of term-weight have been computed using a TWS, the fol-
lowing step is to calculate the similarity matching value between the selected vector and
the web-pages vectors in the web-page collection. Then, the web-pages are retrieved in de-
scending order of their similarity values. The highest ranking web-page will be the most
similar web-page to the selected web-pages. The similarity matching procedure simulates
the automatic system measurement for the relevance levels of the web-pages to the selected
web-pages. The more accurate similarity matching function is the more of effective RSs
accuracy obtained. There are five well-known similarity matching functions that are widely
used in VSM. These functions are inner product, cosine similarity, Dice, Jaccard and Eu-
clidean distance between web-page vectors [21, 15]. Several research has been reported about
these functions in [21]. The Cosine similarity is the most widely used and the most efficient
similarity function in VSM. The cosine similarity function between web-page D and selected

8
web-page Q (Cosine Similarity(D, Q)) is defined by:
Σn Wid · Wiq
Cosine Similarity(D, Q) = q i=1 (5)
Σni=1 Wid2 · Σni=1 Wiq2
In the above equation, n is the number of index terms that exist in the unseen web-page
D and selected web-page Q, Wid is the weight of term i in web-page D and Wiq is the weight
of the same term i in selected web-page Q.

4.1.2. Probabilistic Models (PM)

Probabilistic models in information retrieval are learning approaches based on the bag-
of-words representation in which each document or web-page is a vector of term weights.
Probabilistic models are based on the theory of probability with some statistical basis.
Greengrass [21] illustrated some of the well-known probabilistic IR models such as Okapi.
The next subsection is presenting the Okapi probabilistic model.

Two-Poisson Probabilistic Model (Okapi-Best Match Model)

Robertson et al. [22] proposed a term weighting function based on Two-Poisson distribu-
tion Models. This term weighting function was used in the Okapi IR system at City London
University. It is the most successful term weighting function in TREC (Text Retrieval Con-
ference) competitions. More details of the Okapi weighting function can be described using
the following notations. Let N be the number of web-pages in the collection and ni is the
number of web-pages that contain term i. Moreover, Let Rq be the total number of relevant
web-pages to selected web-page q and ri the number of relevant web-pages that contain
term i. Further, dtfi is the term frequency of term i in the web-page d which has length
dl. The length of the web-page means the total number of index terms frequencies in the
web-page. If the web-page has the average length of the web-pages existing in the web-page
collection, it will be called Avgdl. Using the previous notations, the Okapi weight for term
i in web-page d can be represented by the following equation:
dtfi
w= · w(1) , and (6)
(K1 · dl)/(Avgdl) + dtfi
(ri + 0.5)/(Rq − ri + 0.5)
w(1) = log (7)
(ni − ri + 0.5)/(N − ni − Rq + ri + 0.5)
where K1 represents a constant and usually has a value of either 1 or 2 for TREC document
collections. Robertson et. al used a new similarity matching function in their research [23].
This similarity matching function is called Best Match (BM). In their experimental study,
three formula for Best Match have been used. They are called BM1, BM11 and BM15.
These matching functions represent the Dot Product similarity function [21, 15] between the
web-page d and selected web-page q for a vector of weights (BM15 which is the Dot product
plus correction factor). The equations of the BM25 similarity matching function (BM25
means Best Match version 25) is given by:

9
 X (K1 + 1) · dtfi
BM 25(Q, D) = dl
·
term i ∈q
K1 · ((1 − b) + b · Avgdl ) + dtfi
(ri + 0.5)/(Rq − ri + 0.5) qtfi
log · +
(ni − ri + 0.5)/(N − ni − Rq + ri + 0.5) K3 + qtfi
(Avgdl − dl)
K2 · ql ·
(Avgdl + dl)
(8)

where K1 , K2 , K3 and b are constants that are usually chosen by a trial and error procedure.
The simplest form of BM25 by assigning zero values for Rq and ri is:

X (K1 + 1) · dtfi
BM 25(Q, D) = dl
·
term i ∈q
K1 · ((1 − b) + b · Avgdl ) + dtfi
N − ni − +0.5 qtfi
log · +
ni + 0.5 K3 + qtfi
(Avgdl − dl)
K2 · ql ·
(Avgdl + dl)
(9)

TREC tracks such as TREC-9 [21, 24], have determined the default values for K1 and
b as 1.2 and 0.7 respectively, while K3 is often set to either 0 or 1000 and K2 has often
set to 0. Okapi-BM25 was the best term-weighting function by tuning its constants to
the suitable values based on the web-page collection. However, the need for adjusting the
suitable constants will require the prior knowledge for the relevance feedback of the selected
web-page with the web-page collections regardless of using the relevance judgement in the
BM25 equations. If the constants have zero values in Okapi-BM25, the function will be an
Inner Product similarity function in VSM with Inverse Document Frequency (IDF) term-
weighting scheme.

4.2. Machine Learning

4.2.1. Unsupervised Machine Learning
In this category, unsupervised machine learning techniques are used in recommender
systems. These techniques have been reported in [2]. The most well-known techniques are
various clustering techniques for the data items based on the similarity of their features.
However, the heuristic problems on the data features may not be the best method to have
an accurate clusters which may cause the bias on the recommended items of information to
the users.

10
4.2.2. Supervised Machine Learning
The most common issue in IR and CB-RSs research is ranking the retrieved documents
based on user feedback. In the early research, the MM techniques such as VSM based on
TF-IDF or Okapi-BM25 were used [25]. These models were used to rank the retrieved docu-
ments (web-pages) based on their matching similarity to user selected web-pages. However,
using only one scoring method (Term-Weighting Scheme) was not efficient enough for ef-
fective systems. The reason is that the scoring methods such as Okapi-BM25 are limited
to the relevance feedback in terms of retrieving accurate search results [26, 27, 28]. This
highlights the need for using more than one scoring method for ranking the documents with
respect to the user selected web-page. In addition, the importance of the documents on the
web and the host server, among other desirable features, should be considered to rank the
documents. Recently, Tao Qin et al. [29] proposed a new trend of research into ranking web-
pages by producing LETOR datasets. These datasets are distilled benchmarks from search
engines and from the well-known TREC conference collections. These benchmarks contain
more than one term-weighting scheme (scoring methods) as part of the benchmark features.
They also contain some other features that indicate the importance of the web-page on the
web.

There are three categories of LTR approaches [30]: (1) the pointwise method, (2) the
pairwise method and (3) the listwise method. These categories are based on the loss func-
tion or fitness function measurements. The pointwise approach views each single object
(rated web-page-unseen web-page pair) as the learning instance. Examples of pointwise ap-
proaches are Linear Regression (LR) [31], Boosting [32], Gradient Boosted Regression Trees
(GBRT or MART) [33, 34] and Random Forest (RF) [35]. The pairwise approach views
the pair of objects (two rated web-page-unseen web-page pairs for the same user) as the
learning instance. Examples of the pairwise approaches are RankNET (Rank Neural Net)
[36], RankBoost and SVMRank (Rank Support Vector Machine) [37]. The listwise approach
takes the entire retrieved list of objects (the list of rated web-page-unseen web-page pairs
for each user) as the learning instance. Examples of the listwise approaches are ListNET
(Listwise Neural Net) [38] and RankGPES [39].

Although listwise methods have been shown to perform better regarding accuracy than
point-wise and pair-wise approaches [38], the need to improve the performance of LTR ap-
proaches has motivated researchers to propose hybrid methods as well. For example, Sculley
proposed an approach (CoRR) combining linear regression (point-wise) with support vector
machine (pair-wise) [40]. Two other hybrid approaches are LambdaRank and LambdaMART
which combine pair-wise with list-wise methods [41] also were proposed. LambdaRank is
based on RankNET while LambdaMART is the boosted tree from LambdaRank. Both
LambdaMART and LambdaRank have shown better performance regarding IR accuracy
than the method by Mohan et. al. on the Yahoo! LTR Challenge [42]. Thus, the com-
bination of listwise and pointwise techniques has shown to be promising. Muahmmed and
Carman conducted experiments combining listwise with pointwise Random Forest (Hybrid
RF) showing that their hybrid outperformed other both pointwise and listwise RF in com-
11
putational run-time and accuracy [43].

5. Robust Recommender Systems

In this section, we investigate various aspects that cause bias in RSs decisions by rec-
ommending false information of items or false products to users. Firstly, the Bias-based
various attacks on RSs are discussed and how to overcome these issues through Trust-aware
RSs. However, Trust-aware RSs can not guarantee fair decisions for RSs when information
security of the system is broken by Malwares and spywares. In addition to SQL injections
issues that attacks the RSs databases. Furthermore, some users may not be aware of who
they should trust as happen with teenagers on social media. Teenager users on social media
give their trust by adding unknown friends, while Friends zone on social media consider as
the Trust-aware zone of its RSs. Thus, teenagers may be directed to bias decisions through
their friends on Trust-aware CF-RSs on social media.

5.1. Bias-Based Attacks on the Recommender Systems

The main bias in RSs is caused by various Shilling Attacks [44]. In these attacks, the
individuals or bots copy the same actual user rates using fake accounts and then they use
push or nuke feedback to target the increase or decrease the number of recommendations
for specific items of information. Chirita et al. [45] identified 3 metrics that can be used
to detect the shilling attacks based on users ratings. The first metric is called Number
of Prediction-Difference. This metric checks the differences in prediction changes in the
systems before/after removal a user from the system. The large differences between before
and after removing a user indicates that the RS is biased towards that user. The second
metric is called Standard Deviation in User’s Ratings. In this metric, the difference values
between the user ratings for the items to the average ratings existing in the system. The
difference values are the standard deviation values of the user ratings. These standard
deviation values measure the bias impact caused by that user. The mean of these standard
deviations are called Degree of Agreement with Other Users which is the third metrics for
identifying the bias in RSs. The user that has an impact for bias of the system can be a
fake account user or a Gray Sheep User (GSU) [46], while Gray Sheep Group is the group
of users that have an odd idea about RSs items. To distinguish between relevance feedback
by a GSU and a shilling attack by fake account is very difficult to be identified by the above
metrics. Thus, RSs developers proposed a new perspective for establishing RSs which is
Trust-Aware RSs. In this paradigm, a user can give a trust value for other users to accept
their recommendation. Example of this kind of RS is Epinions and Social Networks. In the
following section, we will talk about this kind of RSs.

5.2. Trust-Aware Recommender Systems

Trust-Aware approach is the state-of-the-art technique used for establishing the rec-
ommendation engine applications to prevent the bias with various attacks using fake user

12
accounts. In Trust-Aware RSs, the recommendation for items of information can be accom-
plished through the trusted users for the target active user or through all trusted users in
the RSs. There are three well-known Trust-Aware techniques that have been used in the re-
search literature which are as follows [47, 14]: 1) Trust-based weighted mean, 2) Trust-based
Collaborative Filtering and 3)TidalTrust. In the following sentences, we will give a summary
about each technique. Assuming that the trust weight between an active user a and a user u
is Wa,u which reflects the degree of trust of an active user a in the rates produced by user u.
We also assume that the rate produced by a user u for item of information i is ru,i . Hence,
the predicted rating target P Ra,i of item i for the active user a using Trust-based weighted
mean technique can be calculated by the following equation [47]:
P
u∈RT Wa,u · ru,i
P Ra,i = P (10)
u∈RT Wa,u

where RT is the set of users that evaluated the item i and these users were rated the
item i and they have trust values in the RS Wa,u which is larger than a given threshold. The
second Trust-Aware technique is the Trust-based Collaborative Filtering technique. The
predicted rating for active user a for item of information i can be given by the following
equation [48]:
P
T Wa,u (ru,i − r¯
u)
P Ra,i = r¯a + u∈RP (11)
u∈RT Wa,u

where r¯a is the mean ratings by user a for other items and r¯u is the mean ratings by
users u for the item of information i. The third Trust-Aware technique is TidalTrust. This
technique is similar to Trust-based weighted mean, but it the set of users that active user
a trusted on them contributed in its calculation. Thus, the trust-weight value between an
active user a and users u in equation 10 can be calculated as follows:
P
v∈W OT + (a) Wa,v · Wv,u
Wa,u = P (12)
v∈W OT + (a) Wa,v

where W OT + (a) is the group of users that user a has trust values with them and it
exceeds than a given threshold, while user u is Friend-Of-Friend user a through user v. Fur-
thermore, Wa,v and Wv,u are the trust weight values between users a with v and v with u
respectively.

On the other hand, there is an issue for the existing of ratings for the most item in the
RSs. This problem makes the previous Trust-Aware techniques limited to the coverage of
explicit trusted user ratings. However, researchers such as in [49, 50] propose that implicit
ratings of item-level and profile-level using CF and Similarity techniques can resolve this
problem. Thus, the user ratings values in the previous techniques can be replaced by the
CF rating values, correlation or similarity value between item of information or between
user profiles. Furthermore, Zeigler and Golbeck [51] investigated the relationship between
the rating similarity of the users with the degree of trust. From their experimental results,
13
they conclude that there is a correlation between the trust-levels and the user-similarity for
rating the RS’s items. Thus, there is a strong argument that the Trust-Aware techniques can
prevent RS bias-based on Shilling and Gray-Sheep attacks. Hence, social networks and well-
known online RS prefer to establish their recommendation based on various Trust-Aware
techniques. However, actual trusted RS users may be varied in their decision from trusted
in some categories to bias in some others. Thus, some social network built their recommen-
dation engines based on Circle-based Trust-aware technique. This technique consider the
level of trust should be categorised to multiple categories. Example for this categories that
user A trusts in user B decisions about movies and he does not trust him in decisions about
clothes, while user A trusts in user C decisions about clothes, but not about movies.

6. Conclusion and Future Work

This literature review introduces various techniques used in RSs. The most popular
methods are CF techniques. They are easier to implement and have lower cost compared to
other techniques. Although, the recent trends for RSs are personalised RSs with supervised
machine learning techniques, these methods can be used for targeting a simulation personal
user feedback to be used to recommend the new items of information. These techniques can
guarantee the serendipity in introducing new items of information in RSs based on personal
user feedback rather than biased CF feedback by group of similar users or similar items of
information. Although, the lack or the size of personal user rates can cause the cold-start
problems in RSs. This inspires the needs for trusted-aware RS technique that can be used
for reducing RS bias based rates and reducing the possibility of cold-start problems with
increasing the possibility of serendipity suggestions. Thus, this technique has been widely
used in social media and some marketing RSs. However, there is no guarantee that Trust-
Aware technique is the only method used in online RSs. This is because there is no research
papers presenting the methods used in online daily life RSs, as they are considered private
assets for the commercial companies.

References
[1] J. S. Breese, D. Heckerman, C. Kadie, Empirical analysis of predictive algorithms for collaborative
filtering, in: Proceedings of the Fourteenth Conference on Uncertainty in Artificial Intelligence, UAI’98,
Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1998, pp. 43–52.
URL http://dl.acm.org/citation.cfm?id=2074094.2074100
[2] F. Ricci, L. Rokach, B. Shapira, P. B. Kantor, Recommender Systems Handbook, 1st Edition, Springer-
Verlag New York, Inc., New York, NY, USA, 2010.
[3] G. Linden, B. Smith, J. York, Amazon.com recommendations: Item-to-item collaborative filtering,
IEEE Internet Computing 7 (1) (2003) 76–80. doi:10.1109/MIC.2003.1167344.
URL http://dx.doi.org/10.1109/MIC.2003.1167344
[4] B. Sarwar, G. Karypis, J. Konstan, J. Riedl, Item-based collaborative filtering recommendation algo-
rithms, in: Proceedings of the 10th International Conference on World Wide Web, WWW ’01, ACM,
New York, NY, USA, 2001, pp. 285–295. doi:10.1145/371920.372071.
URL http://doi.acm.org/10.1145/371920.372071
[5] X. Yang, Y. Guo, Y. Liu, H. Steck, A survey of collaborative filtering based social recommender
systems, Computer Communications 41 (Supplement C) (2014) 1 – 10. doi:https://doi.org/10.
14
 1016/j.comcom.2013.06.009.
URL http://www.sciencedirect.com/science/article/pii/S0140366413001722
[6] X. Yang, Y. Guo, Y. Liu, Bayesian-inference-based recommendation in online social networks, IEEE
Transactions on Parallel and Distributed Systems 24 (4) (2013) 642–651. doi:10.1109/TPDS.2012.192.
[7] M. Jamali, M. Ester, Trustwalker: A random walk model for combining trust-based and item-based
recommendation, in: Proceedings of the 15th ACM SIGKDD International Conference on Knowledge
Discovery and Data Mining, KDD ’09, ACM, New York, NY, USA, 2009, pp. 397–406. doi:10.1145/
1557019.1557067.
URL http://doi.acm.org/10.1145/1557019.1557067
[8] A. Mnih, R. R. Salakhutdinov, Probabilistic matrix factorization, in: J. C. Platt, D. Koller, Y. Singer,
S. T. Roweis (Eds.), Advances in Neural Information Processing Systems 20, Curran Associates, Inc.,
2008, pp. 1257–1264.
URL http://papers.nips.cc/paper/3208-probabilistic-matrix-factorization.pdf
[9] Y. Xin, H. Steck, Multi-value probabilistic matrix factorization for ip-tv recommendations, in: Pro-
ceedings of the Fifth ACM Conference on Recommender Systems, RecSys ’11, ACM, New York, NY,
USA, 2011, pp. 221–228. doi:10.1145/2043932.2043972.
URL http://doi.acm.org/10.1145/2043932.2043972
[10] D. Billsus, M. J. Pazzani, Learning collaborative information filters, in: Proceedings of the Fifteenth
International Conference on Machine Learning, ICML ’98, Morgan Kaufmann Publishers Inc., San
Francisco, CA, USA, 1998, pp. 46–54.
URL http://dl.acm.org/citation.cfm?id=645527.657311
[11] S. Deerwester, S. T. Dumais, G. W. Furnas, T. K. Landauer, R. Harshman, Indexing by latent semantic
analysis, Journal of the American Society for Information Science 41 (6) (1990) 391–407. doi:10.1002/
(SICI)1097-4571(199009)41:6<391::AID-ASI1>3.0.CO;2-9.
URL http://dx.doi.org/10.1002/(SICI)1097-4571(199009)41:6<391::AID-ASI1>3.0.CO;2-9
[12] A. Paterek, Improving regularized singular value decomposition for collaborative filtering, in: Proc.
KDD Cup Workshop at SIGKDD’07, 13th ACM Int. Conf. on Knowledge Discovery and Data Mining,
2007, pp. 39–42.
[13] N. Hug, Surprise, a Python library for recommender systems, http://surpriselib.com (2017).
[14] G. Adomavicius, A. Tuzhilin, Context-aware recommender systems, in: F. Ricci, L. Rokach, B. Shapira,
P. B. Kantor (Eds.), Recommender Systems Handbook, Springer US, Boston, MA, 2011, pp. 217–253.
doi:10.1007/978-0-387-85820-3_7.
URL https://doi.org/10.1007/978-0-387-85820-3_7
[15] R. A. Baeza-Yates, B. A. Ribeiro-Neto, Modern Information Retrieval - the concepts and technology
behind search, Second edition, Pearson Education Ltd., Harlow, England, 2011.
[16] G. Salton, A. Wong, C. S. Yang, A vector space model for automatic indexing, Communication of the
ACM 18 (11) (1975) 613–620. doi:10.1145/361219.361220.
[17] K. S. Jones, A statistical interpretation of term specificity and its application in retrieval, Journal of
Documentation 60 (5) (2004) 493–502. doi:10.1108/00220410410560573.
[18] G. Salton, C. Buckley, Term weighting approaches in automatic text retrieval, Information Processing
and Management: an International Journal 24 (5) (1988) 513–523.
[19] T.-Y. Liu, Learning to Rank for Information Retrieval, Springer Berlin Heidelberg, Berlin, Heidelberg,
2011, Ch. The LETOR Datasets, pp. 133–143. doi:10.1007/978-3-642-14267-3_10.
[20] J. Reed, Y. Jiao, T. Potok, B. Klump, M. Elmore, A. Hurson, Tf-icf: A new term weighting scheme
for clustering dynamic data streams, in: 2006 5th International Conference on Machine Learning and
Applications (ICMLA’06), IEEE, Dec., 2006, pp. 258–263. doi:10.1109/ICMLA.2006.50.
[21] E. Greengrass, Information retrieval : A survey, Tech. rep., University of Maryland, USA (Nov., 2000).
URL http://www.csee.umbc.edu/csee/research/cadip/readings/IR.report.120600.book.pdf
[22] S. E. Robertson, S. Walker, Some simple effective approximations to the 2-poisson model for proba-
bilistic weighted retrieval, in: Proceedings of the 17th Annual International ACM SIGIR Conference
on Research and Development in Information Retrieval, SIGIR ’94, Springer-Verlag New York, Inc.,

15
 New York, NY, USA, 1994, pp. 232–241.
[23] S. E. Robertson, S. Walker, M. Beaulieu, Okapi at trec-7: Automatic ad hoc, filtering vlc and interactive
track, in: E. M. Voorheer, D. K. Harman (Eds.), NIST Special Publication 500-242: The Seventh Text
REtrieval Conference (TREC-7), NIST, Gaithersburg, MD, 1998, pp. 253–264.
[24] S. E. Robertson, S. Walker, Microsoft cambridge at trec-9: Filtering track., in: Text REtrieval Confer-
ence (TREC), 2000.
[25] C. D. Manning, P. Raghavan, H. Schutze, Introduction to Information Retrieval, Cambridge University
Press, New York, NY, USA, 2008.
[26] O. Ibrahim, D. Landa-Silva, Term frequency with average term occurrences for textual information
retrieval, Soft Computing 20 (8) (2016) 3045–3061. doi:10.1007/s00500-015-1935-7.
[27] A. Tonon, G. Demartini, P. Cudr-Mauroux, Pooling-based continuous evaluation of informa-
tion retrieval systems, Information Retrieval Journal 18 (5) (2015) 445–472. doi:10.1007/
s10791-015-9266-y.
[28] J. Urbano, Test collection reliability: a study of bias and robustness to statistical assumptions
via stochastic simulation, Information Retrieval Journal 19 (3) (2016) 313–350. doi:10.1007/
s10791-015-9274-y.
[29] T. Qin, T.-Y. Liu, J. Xu, H. Li, Letor: A benchmark collection for research on learning to rank for
information retrieval, Information Retrieval 13 (4) (2010) 346–374. doi:10.1007/s10791-009-9123-y.
[30] T.-Y. Liu, Learning to rank for information retrieval, Foundations and Trends in Information Retrieval
3 (3) (2009) 225–331.
[31] X. Yan, X. G. Su, Linear Regression Analysis: Theory and Computing, World Scientific Publishing
Co., Inc., River Edge, NJ, USA, 2009.
[32] Y. Freund, R. Iyer, R. E. Schapire, Y. Singer, An efficient boosting algorithm for combining preferences,
Journal of Machine Learning Research 4 (2003) 933–969.
[33] J. H. Friedman, Greedy function approximation: A gradient boosting machine, The Annals of Statistics
29 (5) (2001) 1189–1232.
[34] A. Mohan, Z. Chen, K. Weinberger, Web-search ranking with initialized gradient boosted regression
trees, in: Journal of Machine Learning Research, Workshop and Conference Proceedings, Vol. 14, 2011,
pp. 77–89.
[35] L. Breiman, Random forests, Machine Learning 45 (1) (2001) 5–32. doi:10.1023/A:1010933404324.
[36] C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, G. Hullender, Learning to rank
using gradient descent, in: Proceedings of the 22Nd International Conference on Machine Learning,
ICML ’05, ACM, New York, NY, USA, 2005, pp. 89–96. doi:10.1145/1102351.1102363.
[37] H. Li, Learning to Rank for Information Retrieval and Natural Language Processing, Second Edition,
Morgan and Claypool Publishers, 2014.
[38] Z. Cao, T. Qin, T.-Y. Liu, M.-F. Tsai, H. Li, Learning to rank: from pairwise approach to listwise
approach, in: Proceedings of the 24th international conference on Machine learning, ICML ’07, ACM,
New York, NY, USA, 2007, pp. 129–136. doi:10.1145/1273496.1273513.
[39] M. A. Islam, RankGPES: Learning to rank for information retrieval using a hybrid genetic programming
with evolutionary strategies (2013).
[40] D. Sculley, Combined regression and ranking, in: Proceedings of the 16th ACM SIGKDD International
Conference on Knowledge Discovery and Data Mining, KDD ’10, ACM, New York, NY, USA, 2010,
pp. 979–988. doi:10.1145/1835804.1835928.
[41] C. J. C. Burges, From RankNet to LambdaRank to LambdaMART: An overview, Tech. rep., Microsoft
Research (2010).
[42] O. Chapelle, Y. Chang, Yahoo! learning to rank challenge overview, in: Proceedings of the Yahoo!
Learning to Rank Challenge, held at ICML 2010, Haifa, Israel, June 25, 2010, 2011, pp. 1–24.
[43] M. Ibrahim, M. Carman, Comparing pointwise and listwise objective functions for random-forest-based
learning-to-rank, ACM Transaction of Information System 34 (4) (2016) 20:1–20:38. doi:10.1145/
2866571.
[44] S. K. Lam, J. Riedl, Shilling recommender systems for fun and profit, in: Proceedings of the 13th

16
 International Conference on World Wide Web, WWW ’04, ACM, New York, NY, USA, 2004, pp. 393–
402. doi:10.1145/988672.988726.
URL http://doi.acm.org/10.1145/988672.988726
[45] P.-A. Chirita, W. Nejdl, C. Zamfir, Preventing shilling attacks in online recommender systems, in: Pro-
ceedings of the 7th Annual ACM International Workshop on Web Information and Data Management,
WIDM ’05, ACM, New York, NY, USA, 2005, pp. 67–74. doi:10.1145/1097047.1097061.
URL http://doi.acm.org/10.1145/1097047.1097061
[46] B. Gras, A. Brun, A. Boyer, Identifying grey sheep users in collaborative filtering: A distribution-based
technique, in: Proceedings of the 2016 Conference on User Modeling Adaptation and Personalization,
UMAP ’16, ACM, New York, NY, USA, 2016, pp. 17–26. doi:10.1145/2930238.2930242.
URL http://doi.acm.org/10.1145/2930238.2930242
[47] P. Victor, C. Cornelis, M. D. Cock, Trust Networks for Recommender Systems, 1st Edition, Atlantis
Publishing Corporation, 2011.
[48] C.-N. Ziegler, J. Golbeck, Investigating correlations of trust and interest similarity-do birds of a feather
really flock together? (11 2017).
[49] C.-N. Ziegler, G. Lausen, Analyzing correlation between trust and user similarity in online communities,
in: C. Jensen, S. Poslad, T. Dimitrakos (Eds.), Trust Management: Second International Conference,
iTrust 2004, Oxford, UK, March 29 - April 1, 2004. Proceedings, Springer Berlin Heidelberg, Berlin,
Heidelberg, 2004, pp. 251–265. doi:10.1007/978-3-540-24747-0_19.
URL https://doi.org/10.1007/978-3-540-24747-0_19
[50] X. Ma, H. Lu, Z. Gan, Improving recommendation accuracy by combining trust communities and
collaborative filtering, in: Proceedings of the 23rd ACM International Conference on Conference on
Information and Knowledge Management, CIKM ’14, ACM, New York, NY, USA, 2014, pp. 1951–1954.
doi:10.1145/2661829.2662085.
URL http://doi.acm.org/10.1145/2661829.2662085
[51] C.-N. Ziegler, J. Golbeck, Investigating interactions of trust and interest similarity, Decision Support
Systems 43 (2) (2007) 460 – 475, emerging Issues in Collaborative Commerce. doi:https://doi.org/
10.1016/j.dss.2006.11.003.
URL http://www.sciencedirect.com/science/article/pii/S0167923606001655

17