Lecture Slides for

INTRODUCTION
TO
MACHINE
LEARNING
3RD EDITION
ETHEM ALPAYDIN
© The MIT Press, 2014

[email protected]
http://www.cmpe.boun.edu.tr/~ethem/i2ml3e
CHAPTER 6:

DIMENSIONALITY
REDUCTION
 Why Reduce Dimensionality?
3

 Reduces time complexity: Less computation

 Reduces space complexity: Fewer parameters
 Saves the cost of observing the feature
 Simpler models are more robust on small datasets
 More interpretable; simpler explanation
 Data visualization (structure, groups, outliers, etc) if
plotted in 2 or 3 dimensions
 Feature Selection vs Extraction
4

 Feature selection: Choosing k<d important features,

ignoring the remaining d – k
Subset selection algorithms
 Feature extraction: Project the
original xi , i =1,...,d dimensions to
new k<d dimensions, zj , j =1,...,k
 Subset Selection
5

 There are 2d subsets of d features

 Forward search: Add the best feature at each step
 Set of features F initially Ø.
 At each iteration, find the best new feature
j = argmini E ( F  xi )
 Add xj to F if E ( F  xj ) < E ( F )

 Hill-climbing O(d2) algorithm

 Backward search: Start with all features and remove
one at a time, if possible.
 Floating search (Add k, remove l)
 Iris data: Single feature

Chosen

6
 Iris data: Add one more feature to F4

Chosen

7
 Principal Components Analysis
8

 Find a low-dimensional space such that when x is

projected there, information loss is minimized.
 The projection of x on the direction of w is: z = wTx
 Find w such that Var(z) is maximized
Var(z) = Var(wTx) = E[(wTx – wTμ)2]
= E[(wTx – wTμ)(wTx – wTμ)]
= E[wT(x – μ)(x – μ)Tw]
= wT E[(x – μ)(x –μ)T]w = wT ∑ w
where Var(x)= E[(x – μ)(x –μ)T] = ∑
  Maximize Var(z) subject to ||w||=1
maxw1T w1   w1T w1  1
w1

∑w1 = αw1 that is, w1 is an eigenvector of ∑

Choose the one with the largest eigenvalue for Var(z) to be
max
 Second principal component: Max Var(z2), s.t.,
||w2||=1 and orthogonal to w1
maxwT2 w 2   wT2 w 2  1   wT2 w1  0
w2

∑ w2 = α w2 that is, w2 is another eigenvector of ∑

and so on.
9
 What PCA does
10

z = WT(x – m)
where the columns of W are the eigenvectors of ∑
and m is sample mean
Centers the data at the origin and rotates the axes
 How to choose k ?
11

 Proportion of Variance (PoV) explained

1  2    k
1  2    k    d

when λi are sorted in descending order

 Typically, stop at PoV>0.9
 Scree graph plots of PoV vs k, stop at “elbow”
12
13
 Feature Embedding
14

 When X is the Nxd data matrix,

XTX is the dxd matrix (covariance of features, if mean-
centered)
XXT is the NxN matrix (pairwise similarities of instances)
 PCA uses the eigenvectors of XTX which are d-dim and can
be used for projection
T
 Feature embedding uses the eigenvectors of XX which are
N-dim and which give directly the coordinates after
projection
 Sometimes, we can define pairwise similarities (or distances)
between instances, then we can use feature embedding
without needing to represent instances as vectors.
 Factor Analysis
15

 Find a small number of factors z, which when

combined generate x :
xi – µi = vi1z1 + vi2z2 + ... + vikzk + εi

where zj, j =1,...,k are the latent factors with

E[ zj ]=0, Var(zj)=1, Cov(zi ,, zj)=0, i ≠ j ,
εi are the noise sources
E[ εi ]= ψi, Cov(εi , εj) =0, i ≠ j, Cov(εi , zj) =0 ,
and vij are the factor loadings
 PCA vs FA
16

 PCA From x to z z = WT(x – µ)

 FA From z to x x – µ = Vz + ε

x z

z x
 Factor Analysis
17

 In FA, factors zj are stretched, rotated and

translated to generate x
 Singular Value Decomposition and
18
Matrix Factorization
 Singular value decomposition: X=VAWT
V is NxN and contains the eigenvectors of XXT
W is dxd and contains the eigenvectors of XTX
and A is Nxd and contains singular values on its first
k diagonal
 X=u1a1v1T+...+ukakvkT where k is the rank of X
 Matrix Factorization
19

 Matrix factorization: X=FG

F is Nxk and G is kxd

Latent semantic indexing

 Multidimensional Scaling
20

 Given pairwise distances between N points,

dij, i,j =1,...,N
place on a low-dim map s.t. distances are preserved
(by feature embedding)
 z = g (x | θ ) Find θ that min Sammon stress

E  | X   
z r
z  x x
s r s

2

s 2
r ,s x xr


 gx |  gx |   x
r s r
x s

2

s 2
r ,s x x
r
 Map of Europe by MDS
21

Map from CIA – The World Factbook: http://www.cia.gov/

Linear Discriminant Analysis

 Find a low-dimensional
space such that when x
is projected, classes are
well-separated.
 Find w that maximizes

J w  
m1  m2  2

s1  s2
2 2

m1 
t x r
w T t t

s  t w x  m1  r
2 T t 2 t

r t 1
t
22
 Between-class scatter:
m1  m2   w m1  w m 2 
2 T T 2

 w T m1  m 2 m1  m 2 T w
 w T SB w where SB  m1  m 2 m1  m 2 T

 Within-class scatter:
s  t w x  m1  r
2 T t 2 t
1

 t w x  m1 x  m1  wr t  w T S1w
T t t T

where S1  t x  m1 x  m1  r
t t T t

s12  s12  w T SW w where SW  S1  S 2

23
 Fisher’s Linear Discriminant
24

 Find w that max

w SB w w m1  m 2 
T 2
T
Jw   T 
w SW w w T SW w
LDA soln:
w  c  SW1 m1  m2 


 Parametric soln:
w   1 μ1  μ 2 
when px|C i  ~ N μ i ,  
 K>2 Classes
25

 Within-class scatter:
Si  t ri x  m i x  m i 
K
SW   Si t t t T

i 1

 Between-class scatter:
K
1 K
SB   Ni m i  m m i  m  T
m  mi
i 1 K i 1
 Find W that max JW   WT SB W
WT SW W
The largest eigenvectors of SW-1SB; maximum rank of K-1
26
 PCA vs LDA
27
 Canonical Correlation Analysis
28

 X={xt,yt}t ; two sets of variables x and y x

 We want to find two projections w and v st when x
is projected along w and y is projected along v, the
correlation is maximized:
 CCA
29

 x and y may be two different views or modalities;

e.g., image and word tags, and CCA does a joint
mapping
 Isomap
30

 Geodesic distance is the distance along the

manifold that the data lies in, as opposed to the
Euclidean distance in the input space
 Isomap
31

 Instances r and s are connected in the graph if

||xr-xs||<e or if xs is one of the k neighbors of xr
The edge length is ||xr-xs||
 For two nodes r and s not connected, the distance is
equal to the shortest path between them
 Once the NxN distance matrix is thus formed, use
MDS to find a lower-dimensional mapping
 Optdigits after Isomap (with neighborhood graph).
150

100 2
22222
2
2
50 33 22 2
7 7777 1 11 313
3
3
77 7 7 7 4 111 1
1 3 338
1 8 83
7 44999 5 5 5 98 38
0 9 4
994949 5 9 88
4 0
88 0 0
0 00
-50 000
6
4 6 66 0
6 66
4
44
-100
4

-150
-150 -100 -50 0 50 100 150

Matlab source from http://web.mit.edu/cocosci/isomap/isomap.html

32
 Locally Linear Embedding
33

1. Given xr find its neighbors xs(r)

2. Find Wrs that minimize
2

E (W| X )   x r   Wrs x(sr )

r s

3. Find the new coordinates zr that minimize

2

E (z | W)   z r   Wrs z(sr )
r s
34
 LLE on Optdigits
35

0 000

7
7777
7
6 666 7
7 9 9
1 66 399 47
84 4
8383
9334
957
9
44
389 93
41
9
8 34
3
484 1
1
4 4 1 82 282
1 1 22 222
9
8 25
1
1
1 55

5
1

1
-3.5 -3 -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5

Matlab source from http://www.cs.toronto.edu/~roweis/lle/code.html

 Laplacian Eigenmaps
36

 Let r and s be two instances and Brs is their similarity, we

want to find zr and zs that

 Brs can be defined in terms of similarity in an original

space: 0 if xr and xs are too far, otherwise

 Defines a graph Laplacian, and feature embedding

returns zr
 Laplacian Eigenmaps on Iris
37

Spectral clustering (chapter 7)