Support Vector

Machines

MEDINFO 2004,
T02: Machine Learning Methods for Decision Support and Discovery
Constantin F. Aliferis & Ioannis Tsamardinos
Discovery Systems Laboratory
Department of Biomedical Informatics
Vanderbilt University

1
Support Vector Machines
 Decision surface is a hyperplane (line in 2D) in
feature space (similar to the Perceptron)
 Arguably, the most important recent discovery in
machine learning
 In a nutshell:
 map the data to a predetermined very high-
dimensional space via a kernel function
 Find the hyperplane that maximizes the margin
between the two classes
 If data are not separable find the hyperplane that
maximizes the margin and minimizes the (a weighted
average of the) misclassifications

2
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
3
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
4
 Which Separating Hyperplane to Use?
Var1

Var2 5
 Maximizing the Margin
Var1 IDEA 1: Select the
separating
hyperplane that
maximizes the
margin!

Margin
Width

Margin
Width
Var2 6
 Support Vectors
Var1

Support Vectors

Margin
Width
Var2 7
 Setting Up the Optimization Problem
Var1
The width of the
margin is:
2k
w
 
 w x b  k So, the problem is:
w
2k
max
  w
w  x  b  k
k k Var2
s.t. (w  x  b)  k , x of class 1
  (w  x  b)   k , x of class 2
w x  b  0

8
 Setting Up the Optimization Problem
Var1
There is a scale and
unit for data so that
k=1. Then problem
becomes:

2
  max
 w x  b 1 w
w s.t. (w  x  b)  1, x of class 1
  (w  x  b)  1, x of class 2
w  x  b  1
1 1 Var2
 
w x  b  0

9
Setting Up the Optimization Problem
 If class 1 corresponds to 1 and class 2 corresponds
to -1, we can rewrite
(w  xi  b)  1, xi with yi  1
(w  xi  b)  1, xi with yi  1
 as
yi (w  xi  b)  1, xi
 So the problem becomes:
2 1 2
max min w
w or 2
s.t. yi (w  xi  b)  1, xi s.t. yi (w  xi  b)  1, xi

10
Linear, Hard-Margin SVM
Formulation
 Find w,b that solves

1 2
min w
2
s.t. yi (w  xi  b)  1, xi
 Problem is convex so, there is a unique global minimum value
(when feasible)
 There is also a unique minimizer, i.e. weight and b value that
provides the minimum
 Non-solvable if the data is not linearly separable
 Quadratic Programming
 Very efficient computationally with modern
constraint optimization engines (handles
thousands of constraints and training instances).
11
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
12
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
13
 Non-Linearly Separable Data

Var1 Introduce slack

i variables i

Allow some
instances to fall
i within the margin,
  but penalize them
 w x  b 1
w
 
w  x  b  1
1 1 Var2
 
w x b  0
14
 Formulating the Optimization Problem
Constraint becomes :

Var1 yi (w  xi  b)  1  i , xi
i
i  0

Objective function
penalizes for
i
misclassified instances
  and those within the
 w x  b 1
w margin
1
min w  C  i
2
 
w  x  b  1
1 1 Var2
2 i
 
w x b  0
C trades-off margin width
and misclassifications 15
Linear, Soft-Margin SVMs
1
min w  C  i yi (w  xi  b)  1  i , xi
2

2 i i  0
 Algorithm tries to maintain i to zero while maximizing
margin
 Notice: algorithm does not minimize the number of
misclassifications (NP-complete problem) but the sum of
distances from the margin hyperplanes
 Other formulations use i2 instead
 As C, we get closer to the hard-margin solution

16
 Robustness of Soft vs Hard Margin
SVMs

Var1 Var1

i

i

Var2
 
w x  b  0 Var2
 
w x  b  0

Soft Margin SVN Hard Margin SVN

17
Soft vs Hard Margin SVM
 Soft-Margin always have a solution
 Soft-Margin is more robust to outliers
 Smoother surfaces (in the non-linear case)
 Hard-Margin does not require to guess the
cost parameter (requires no parameters at
all)

18
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
19
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
20
 Disadvantages of Linear Decision
Surfaces
Var1

Var221
 Advantages of Non-Linear Surfaces
Var1

Var222
 Linear Classifiers in High-
Dimensional Spaces
Constructed
Var1
Feature 2

Var2
Constructed
Feature 1
Find function (x) to map to
a different space
23
Mapping Data to a High-Dimensional
Space
• Find function (x) to map to a different space, then
SVM formulation becomes:
1 s.t. yi ( w   ( x )  b)  1   i , xi
min w  C  i
2

2 i i  0

• Data appear as (x), weights w are now weights in

the new space
• Explicit mapping expensive if (x) is very high
dimensional
• Solving the problem without explicitly mapping the
data is desirable
24
The Dual of the SVM Formulation
 Original SVM formulation
1 2
 n inequality constraints
min w  C i
 n positivity constraints w ,b 2 i
 n number of  variables

s.t. yi ( w   ( x )  b)  1   i , xi
i  0
 The (Wolfe) dual of this
1
problem
min   i j yi y j ( ( xi )   ( x j ))    i
 one equality constraint ai 2
i, j i
 n positivity constraints
 n number of  variables s.t. C   i  0, xi
(Lagrange multipliers)
 Objective function more
complicated
 y
i
i i 0

 NOTICE: Data only appear

as (xi)  (xj)
25
The Kernel Trick
 (xi)  (xj): means, map data into new space, then take the
inner product of the new vectors
 We can find a function such that: K(xi  xj) = (xi)  (xj), i.e., the
image of the inner product of the data is the inner product of the
images of the data
 Then, we do not need to explicitly map the data into the high-
dimensional space to solve the optimization problem (for
training)
 How do we classify without explicitly mapping the new
instances? Turns out
sgn( wx  b)  sgn(  i yi K ( xi , x )  b)
i

where b solves  j ( y j   i yi K ( xi , x j )  b  1)  0,
i

for any j with  j  0

26
Examples of Kernels
 Assume we measure two quantities, e.g. expression
level of genes TrkC and SonicHedghog (SH) and we
use the mapping:
 : xTrkC , x SH  {x TrkC
2 2
, x SH , 2 xTrkC x SH , xTrkC , x SH ,1}
 Consider the function:
K ( x  z )  ( x  z  1) 2
 We can verify that:
( x)  ( z ) 
2
x TrkC 2
z TrkC  x SH
2 2
z SH  2 xTrkC x SH zTrkC z SH  xTrkC zTrkC  x SH z SH  1 
 ( xTrkC zTrkC  x SH z SH  1) 2  ( x  z  1) 2  K ( x  z )

27
Polynomial and Gaussian Kernels
K ( x  z )  ( x  z  1) p
 is called the polynomial kernel of degree p.
 For p=2, if we measure 7,000 genes using the kernel once
means calculating a summation product with 7,000 terms then
taking the square of this number
 Mapping explicitly to the high-dimensional space means
calculating approximately 50,000,000 new features for both
training instances, then taking the inner product of that (another
50,000,000 terms to sum)
 In general, using the Kernel trick provides huge computational
savings over explicit mapping!
 Another commonly used Kernel is the Gaussian (maps to a
dimensional space with number of dimensions equal to the
number of training cases):
K ( x  z )  exp( x  z / 2 )2

28
The Mercer Condition
 Is there a mapping (x) for any symmetric
function K(x,z)? No
 The SVM dual formulation requires
calculation K(xi , xj) for each pair of training
instances. The array Gij = K(xi , xj) is called
the Gram matrix
 There is a feature space (x) when the
Kernel is such that G is always semi-positive
definite (Mercer condition)

29
Support Vector Machines
 Three main ideas:
1. Define what an optimal hyperplane is (in way
that can be identified in a computationally
efficient way): maximize margin
2. Extend the above definition for non-linearly
separable problems: have a penalty term for
misclassifications
3. Map data to high dimensional space where it
is easier to classify with linear decision
surfaces: reformulate problem so that data is
mapped implicitly to this space
30
Other Types of Kernel Methods
 SVMs that perform regression
 SVMs that perform clustering
 -Support Vector Machines: maximize margin while
bounding the number of margin errors
 Leave One Out Machines: minimize the bound of the
leave-one-out error
 SVM formulations that take into consideration
difference in cost of misclassification for the different
classes
 Kernels suitable for sequences of strings, or other
specialized kernels
31
Variable Selection with SVMs
 Recursive Feature Elimination
 Train a linear SVM
 Remove the variables with the lowest weights (those
variables affect classification the least), e.g., remove
the lowest 50% of variables
 Retrain the SVM with remaining variables and repeat
until classification is reduced
 Very successful
 Other formulations exist where minimizing the number
of variables is folded into the optimization problem
 Similar algorithm exist for non-linear SVMs
 Some of the best and most efficient variable selection
methods

32
Comparison with Neural Networks
Neural Networks SVMs
 Hidden Layers map to lower  Kernel maps to a very-high
dimensional spaces dimensional space
 Search space has multiple  Search space has a unique
local minima minimum
 Training is expensive  Training is extremely
 Classification extremely efficient
efficient  Classification extremely
 Requires number of hidden efficient
units and layers  Kernel and cost the two
 Very good accuracy in parameters to select
typical domains  Very good accuracy in
typical domains
 Extremely robust

33
Why do SVMs Generalize?
 Even though they map to a very high-
dimensional space
 They have a very strong bias in that space
 The solution has to be a linear combination of
the training instances
 Large theory on Structural Risk Minimization
providing bounds on the error of an SVM
 Typically the error bounds too loose to be of
practical use

34
MultiClass SVMs
 One-versus-all
Train n binary classifiers, one for each class against all other
classes.
 Predicted class is the class of the most confident classifier
 One-versus-one
 Train n(n-1)/2 classifiers, each discriminating between a pair
of classes
 Several strategies for selecting the final classification based
on the output of the binary SVMs
 Truly MultiClass SVMs
 Generalize the SVM formulation to multiple categories
 More on that in the nominated for the student paper award:
“Methods for Multi-Category Cancer Diagnosis from Gene
Expression Data: A Comprehensive Evaluation to Inform
Decision Support System Development”, Alexander Statnikov,
Constantin F. Aliferis, Ioannis Tsamardinos

35
Conclusions
 SVMs express learning as a mathematical
program taking advantage of the rich theory
in optimization
 SVM uses the kernel trick to map indirectly to
extremely high dimensional spaces
 SVMs extremely successful, robust, efficient,
and versatile while there are good theoretical
indications as to why they generalize well

36
37