SVD and PCA
Real data usually have thousands, or millions of
dimensions
E.g., web documents, where the dimensionality is the
vocabulary of words
Facebook graph, where the dimensionality is the
number of users
Huge number of dimensions causes problems
Data becomes very sparse, some algorithms become
meaningless (e.g. density based clustering)
The complexity of several algorithms depends on the
dimensionality and they become infeasible.
Usually the data can be described with fewer
dimensions, without losing much of the meaning
of the data.
The data reside in a space of lower dimensionality
Essentially, we assume that some of the data is
noise, and we can approximate the useful part
with a lower dimensionality space.
Dimensionality reduction does not just reduce the
amount of data, it often brings out the useful part of
the data
We have already seen a form of
dimensionality reduction
LSH, and random projections reduce the
dimension while preserving the distances
SVD is “the Rolls-Royce and the Swiss Army
Knife of Numerical Linear Algebra.”*
*Dianne O’Leary, MMDS ’06
We are given n objects and d attributes describing the
objects. Each object has d numeric values describing
it.
We will represent the data as a n×d real matrix A.
We can now use tools from linear algebra to process the
data matrix
Our goal is to produce a new n×k matrix B such that
It preserves as much of the information in the original
matrix A as possible
It reveals something about the structure of the data in A
d terms
(e.g., theorem, proof, etc.)
n documents
Aij = frequency of the j-th
term in the i-th document
Find subsets of terms that bring documents
together
d movies
n customers
Aij = rating of j-th
product by the i-th
customer
Find subsets of movies that capture the behavior or
the customers
We assume that vectors are column vectors.
We use for the transpose of vector (row vector)
Dot product: (1×, ×1 → 1×1)
The dot product is the projection of vector on (and vice versa)
4
1, 2, 3 1 12
2
cos,
If |||| 1 (unit vector) then is the projection length of on
4
1, 2, 3 1 0 orthogonal vectors
2
Orthonormal vectors: two unit vectors that are orthogonal
An n×m matrix A is a collection of n row vectors and m column
vectors
| | |
| | |
Matrix-vector multiplication
Right multiplication : projection of u onto the row vectors of , or
projection of row vectors of onto .
Left-multiplication : projection of onto the column vectors of ,
or projection of column vectors of onto
Example:
1 0
1,2,3 0 1 1,2
0 0
Row space of A: The set of vectors that can be written
as a linear combination of the rows of A
All vectors of the form
Column space of A: The set of vectors that can be
written as a linear combination of the columns of A
All vectors of the form .
Rank of A: the number of linearly independent row (or
column) vectors
These vectors define a basis for the row (or column) space
of A
In a rank-1 matrix, all columns (or rows) are
multiples of the same column (or row) vector
1 2 1
2 4 2
3 6 3
All rows are multiples of 1,2, 1
All columns are multiples of 1,2,3
External product: (×1 , 1×! → ×!)
The resulting ×! has rank 1: all rows (or columns)
are linearly dependent
(Right) Eigenvector of matrix A: a vector v
such that "
": eigenvalue of eigenvector
A square matrix A of rank r, has r orthonormal
eigenvectors , , … , $ with eigenvalues
" , " , … , "$ .
Eigenvectors define an orthonormal basis for
the column space of A
)
0
)
% Σ ' , , ⋯ , $
⋱ ⋮
0
)$ $
[n×m] = [n×r] [r×r] [r×m]
r: rank of matrix A
) , , ) , ⋯ , )$ : singular values of matrix (also, the square roots of
eigenvalues of and )
, , … , $ : left singular vectors of (also eigenvectors of )
, , … , $ : right singular vectors of (also, eigenvectors of )
) - ) - ⋯ - )$ $ $
Special case: A is symmetric positive definite
matrix
" - " - ⋯ - "$ $ $
" , " , ⋯ , "$ , 0: Eigenvalues of A
, , … , $ : Eigenvectors of A
The left singular vectors are an orthonormal
basis for the row space of A.
The right singular vectors are an orthonormal
basis for the column space of A.
If A has rank r, then A can be written as the sum
of r rank-1 matrices
There are r “linear components” (trends) in A.
Linear trend: the tendency of the row vectors of A to align with vector
v
Strength of the i-th linear trend: ||./ || 0/
Document-term matrix
Blue and Red rows (colums) are linearly dependent
A=
There are two prototype documents (vectors of words): blue and red
To describe the data is enough to describe the two prototypes, and the
projection weights for each row
A is a rank-2 matrix
2
1 , 1
2
Document-term matrix
A=
There are two prototype documents and words but
they are noisy
We now have more than two singular vectors, but the
strongest ones are still about the two types.
By keeping the two strongest singular vectors we obtain
most of the information in the data.
▪ This is a rank-2 approximation of the matrix A
nxd nxk kxk kxd
Uk (Vk): orthogonal matrix containing the top k left (right)
singular vectors of A.
Σk: diagonal matrix containing the top k singular values of A
Ak is an approximation of A
Ak is the best approximation of A
The rank-k approximation matrix 3
produced by the top-k singular vectors of A
minimizes the Frobenious norm of the
difference with the matrix A
3 arg max > ?
9:$;<3 9 =3
> ? @ AB >AB
A,B
We can project the row (and column) vectors
of the matrix A into a k-dimensional space
and preserve most of the information
(Ideally) The k dimensions reveal latent
features/aspects/topics of the term
(document) space.
(Ideally) The 3 approximation of matrix A,
contains all the useful information, and what
is discarded is noise
Rows (columns) are linear combinations of k
latent factors
E.g., in our extreme document example there are
two factors
Some noise is added to this rank-k matrix
resulting in higher rank
SVD retrieves the latent factors (hopefully).
A = U Σ VT
features
significant sig. significant
noise noise
= noise
objects
Data: Users rating movies
Sparse and often noisy
Assumption: There are k basic user profiles, and each user
is a linear combination of these profiles
E.g., action, comedy, drama, romance
Each user is a weighted cobination of these profiles
The “true” matrix has rank k
What we observe is a noisy, and incomplete version of this
matrix C
The rank-k approximation C3 is provably close to 3
Algorithm: compute C3 and predict for user and movie
!, the value C3 !, .
Model-based collaborative filtering
PCA is a special case of SVD on the centered
covariance matrix.
Goal: reduce the dimensionality while preserving the
“information in the data”
Information in the data: variability in the data
We measure variability using the covariance matrix.
Sample covariance of variables X and Y
@ DA EF GA EH
A
Given matrix A, remove the mean of each column
from the column vectors to get the centered matrix C
The matrix ' I I is the covariance matrix of the
row vectors of A.
We will project the rows of matrix A into a new
set of attributes (dimensions) such that:
The attributes have zero covariance to each other
(they are orthogonal)
Each attribute captures the most remaining variance
in the data, while orthogonal to the existing attributes
▪ The first attribute should capture the most variance in the
data
For matrix C, the variance of the rows of C when
projected to vector x is given by ) ID
The right singular vector of C maximizes ) !
Input: 2-d dimensional points
5 Output:
2nd (right)
singular 1st (right) singular vector:
vector direction of maximal variance,
4
2nd (right) singular vector:
direction of maximal variance,
3 after removing the projection of
the data along the first singular
1st (right) singular vector.
vector
2
4.0 4.5 5.0 5.5 6.0
5
2nd (right)
singular σ1: measures how much of the
4
vector data variance is explained by the
first singular vector.
σ2: measures how much of the
3 data variance is explained by the
σ1 second singular vector.
1st (right) singular
vector
2
4.0 4.5 5.0 5.5 6.0
The variance in the direction of the k-th principal component
is given by the corresponding singular value σk2
Singular values can be used to estimate how many
components to keep
Rule of thumb: keep enough to explain 85% of the variation:
k
∑j =1
σ 2
j
n
≈ 0 . 85
∑j =1
σ 2
j
drugs
⋯ <
⋮ ⋱ ⋮ students
< ⋯ <<
legal illegal
AB : usage of student i of drug j
%Σ'
First right singular vector
Drug 1
More or less same weight to all drugs
Discriminates heavy from light users
Second right singular vector
Positive values for legal drugs, negative for illegal
Drug 2
The chosen vectors are such that minimize the sum of square differences
between the data vectors and the low-dimensional projections
1st (right) singular
vector
2
4.0 4.5 5.0 5.5 6.0
Latent Semantic Indexing (LSI):
Apply PCA on the document-term matrix, and
index the k-dimensional vectors
When a query comes, project it onto the k-
dimensional space and compute cosine similarity
in this space
Principal components capture main topics, and
enrich the document representation
# SVD
dat = seq(1,240,2)
X = matrix(dat,ncol=12)
s = svd(X)
A = diag(s$d)
s$u %*% A %*% t(s$v) # X = U A V'
dat = seq(1,240,2)
X = matrix(dat,ncol=12)
s = svd(X, nu = nrow(X), nv = ncol(X))
A = diag(s$d)
A = cbind(A, 0) # Add two columns with zero, in order to A have the same dimensions of X.
A = cbind(A, 0)p
s$u %*% A %*% t(s$v) # X = U A V'
install.packages("jpeg")
library(jpeg)
tux = readJPEG("tux.jpg")
tux = imagematrix(tux,type='grey')
plot(tux)
reduce <- function(A,dim) {
#Calculates the SVDprincomp
sing <- svd(A)
#Approximate each result of SVD with the given dimension
u<-as.matrix(sing$u[, 1:dim])
v<-as.matrix(sing$v[, 1:dim])
d<-as.matrix(diag(sing$d)[1:dim, 1:dim])
#Create the new approximated matrix
return(imagematrix(u%*%d%*%t(v),type='grey'))
}
tux_d = svd(tux)
length(tux_d$d)
plot(reduce(tux,1))
# 90% reduction
plot(reduce(tux,35))
plot(pc$scores[,2], pc$scores[,1])
# PCA
pc = princomp(iris2)
summary(pc)
pc$scores
pc$loadings
