Support Vector Machines -

an Introduction

Ron Meir

Department of Electrical Engineering

Technion, Israel
Support Vector Machines - An Introduction

Sources of Information

Web http://www.kernel-machines.org/

Tutorial C.J.C. Burges. A Tutorial on Support

Vector Machines for Pattern Recognition
(download from above site)

Books

? V. Vapnik, Statistical Learning Theory,

1998.
? N. Cristianini and J. Shawe-Taylor, An
Introduction to Support Vector Machines,
2000.
? A. Smola and B. Schölkopf, Learning with
Kernels, 2002.

Optimization book D. Bertsekas, Nonlinear

Programming, Second Edition 1999.

June 2002 2/44

Support Vector Machines - An Introduction

Application Domains

Supervised Learning

? Pattern Recognition - state of the art

results for OCR, text classification,
Biological sequencing
? Regression and time series - good results

Unsupervised Learning

? Dimensionality Reduction - Non linear

principal component analysis
? Clustering
? Novelty detection

Reinforcement Learning: Some preliminary

results

June 2002 3/44

Support Vector Machines - An Introduction

Classification I

3 1 3 ?

The problem:

Input: x feature vector

Label: y ∈ {1, 2, . . . , k}
Data: {(xi , yi )}m
i=1

Unknown source: x ∼ p(x)

Target: y = f (x)

Objective: Given new x, predict y so that

probability of error is minimal

June 2002 4/44

Support Vector Machines - An Introduction

Classification II
The ‘Model’

Hypothesis class: H : Rd 7→ {±1}

Loss: `(y, h(x)) = I[y 6= h(x)])
Generalization: L(h) = E{`(Y, h(X))}
Objective: Find h ∈ H which mini-
mizes L(h)

Caveat: Only have data at our disposal

‘Solution’: Form empirical estimator which
‘generalizes well’
Question: How can we efficiently construct
complex hypotheses with good gen-
eralization?
Focus: Two-class problem, y ∈ {−1, +1}

June 2002 5/44

Support Vector Machines - An Introduction

Linearly Separable Classes


 +1 w> x + b > 0
Y = sgn[w> x + b > 0] =
 −1 w> x + b ≤ 0

Problem: Many solutions! Some are very

poor
Task: Based on data, select hyper-plane
which works well ‘in general’

June 2002 6/44

Support Vector Machines - An Introduction

Selection of a Good Hyper-Plane

Objective: Select a ‘good’ hyper-plane using

only the data!
Intuition: (Vapnik 1965) - assuming linear
separability

(i) Separate the data

(ii) Place hyper-plane ‘far’ from

data

June 2002 7/44

Support Vector Machines - An Introduction

VC Dimension

Given: H = {h : Rd 7→ {−1, +1}}

Question: How complex is the class?
§

Shattering: H shatters a set X if H

achieves all dichotomies on X

All 8 dichotomies 14 dichotomies

VCdim=3
6 dichotomies

VC-dimension The size of the largest shat-

tered subset of X
Hyper-planes VCdim(H) = d + 1

June 2002 8/44

Support Vector Machines - An Introduction

What is the True Performance?

For h ∈ H
L(h) Probability of miss-classification
L̂n (h) Empirical fraction of miss-classifications
Vapnik and Chervonenkis 1971: For any
distribution with prob. 1 − δ, ∀h ∈ H,
s
VCdim(H) log n + log 1δ
L(h) < L̂n (h) + c
| {z } n
emp. error | {z }
complexity penalty

^
L ( h ) L( h )
n

June 2002 9/44

Support Vector Machines - An Introduction

An Improved VC Bound I
Hyper-plane: H(w, b) = {x : w> x + b = 0}
Distance of a point from a hyper-plane:

w> x + b
d(x, H(w, b)) =
kwk

d(x,H(w,b))

Optimal hyper-plane (linearly separable case)

max min d(xi , H(w, b))

w,b 1≤i≤n

June 2002 10/44

Support Vector Machines - An Introduction

An Improved VC Bound II
Canonical hyper-plane:

min |w> xi + b| = 1
1≤i≤n

(No loss of generality)

Improved VC Bound (Vapnik 95) VC
dimension of set of canonical hyper-planes such
that

kwk ≤ A
xi ∈ Ball of radius L

is
VCdim ≤ min(A2 L2 , d) + 1

Observe: Constraints reduce VC-dim bound

Canonical hyper-planes with mini-
mal norm yields best bound
Suggestion: Use hyper-plane with minimal
norm

June 2002 11/44

Support Vector Machines - An Introduction

The Optimization Problem I

Canonical hyper-planes: (xi , w ∈ Rd )

min |w> xi + b| ≥ 1
1≤i≤n

Support vectors:

{xi : |w> xi + b| = 1}

Margin

Margin Distance between hyper-planes defined

by support vectors

June 2002 12/44

Support Vector Machines - An Introduction

The Optimization Problem II

Distance from support vector to H(w, b)

w > xi + b ±1
=
kwk kwk

¯ ¯
¯ 1 −1 ¯ 2
Margin = ¯
¯ − ¯ =
kwk kwk ¯ kwk

1 >
minimize w w
2
subject to yi (w> xi + b) ≥ 1 i = 1, 2, . . . , n

1. Convex quadratic program

2. Linear inequality constraints (many!)
3. d + 1 parameters, n constraints

June 2002 13/44

Support Vector Machines - An Introduction

Convex Optimization
Problem:

minimize f (x)
subject to hi (x) = 0, i = 1, . . . , m,
gj (x) ≤ 0, j = 1, . . . , r

Active constraints: A(x) = {j : gj (x)=0}

Lagrangian
m
X r
X
L(x, λ, µ) = f (x) + λi hi (x) + µj hj (x)
i=1 j=1

Sufficient conditions for minimum: (KKT)

Let x∗ be a local minimum. Then ∃λ∗ , µ∗ s.t.

∇x L(x∗ , λ∗ , µ∗ ) = 0
µ∗j ≥ 0 j = 1, 2, . . . , r
µ∗j = 0 / A(x∗ )
∀j ∈

June 2002 14/44

Support Vector Machines - An Introduction

The Dual Problem I

Motivation:

? Many inequality constraints

? High (sometimes infinite) input dimension
Primal Problem

minimize f (x)
subject to e>
i x = di , i = 1, . . . , m,
a>
j x ≤ bj , j = 1, . . . , r

Lagrangian
m
X r
X
LP (x, λ, µ) = f (x)+ λi (e>
i x−di )+ µj (a>
j x−bj )
i=1 j=1

June 2002 15/44

Support Vector Machines - An Introduction

The Dual Problem II

Dual Lagrangian

LD (λ, µ) = inf LP (x, λ, µ)

x

Dual Problem

maximize LD (λ, µ)
λ,µ

subject to µ ≥ 0

Observation:
? LP (x, λ, µ) quadratic ⇒ LD (λ, µ) quadratic
? Constraints in Dual greatly simplified
? m + r variables, r constraints

Duality Theorem:
Optimal solutions of P and D coincide

June 2002 16/44

Support Vector Machines - An Introduction

SVM in the Primal Space I

1
minimize kwk2
w,b 2
¡ > ¢
subject to yi w xi + b ≥ 1, i = 1, . . . , n.
n
1 X
LP (w, b, α) = kwk2 − αi [yi (w> xi + b) − 1],
2 i=1

Solution:
n
X
w= α i yi x i
i=1
n
X
0= α i yi (αi ≥ 0)
i=1

KKT condition:

αi = 0 unless yi (w> xi + b) = 1

Sparsity: Often many αi vanish!

June 2002 17/44

Support Vector Machines - An Introduction

SVM in the Primal Space II

Recall
n
X
w= α i yi x i
i=1

Margin

Support vectors: All xi for which αi > 0

Occurs if constraint is
obeyed with equality

June 2002 18/44

Support Vector Machines - An Introduction

Support Vectors in the Dual Space

n
1X
X
max. LD (α) = αi − α i α j yi yj x >
i xj
i=1
2 i,j
n
X
s.t. α i yi = 0 ; αi ≥ 0
i=1

Determination of b: For support vectors

yi (w> xi + b) = 1

Thus
µ ¶
1
b∗ = − min {w∗T xi } + max {w∗T xi }
2 yi =+1 yi =−1

P
Classifier: (Recall w = i α i yi x i )
à n
!
X
f (x) = sgn αi∗ yi x>
i x+b
∗

i=1

June 2002 19/44

Support Vector Machines - An Introduction

Non-Separable Case I

Objective: find a good separating hyper-plane

for the non-separable case
Problem: Cannot satisfy yi [w> xi + b] ≥ 1 for
all i
Solution: Slack variables

w> xi + b ≥ +1 − ξi for yi = +1,

w> xi + b ≤ −1 + ξi for yi = −1,
ξi ≥ 0 k = 1, 2, . . . , n.

An error occurs if ξi > 1. Thus,

n
X
I(ξi > 1) = # errors
i=1

June 2002 20/44

Support Vector Machines - An Introduction

Non-Separable Case II
Proposed solution: Minimize
n
1 2
X
kwk + C I(ξi > 1) (non − convex!)
2 i=1

Suggestion: Replace I(ξi > 1) by ξi (upper

bound)

n
1 2
X
minimize LP (w, ξ) = kwk + C ξi
w,b,ξ 2 i=1

subject to yi (w> xi + b) ≥ 1 − ξi
ξi ≥ 0

Tradeoff: Large C - penalize errors, Small C

penalize complexity

Dual Problem: Same as in separable case,

except that 0 ≤ αi ≤ C

Support vectors: αi > 0 - but lose geometric

interpretation!

June 2002 21/44

Support Vector Machines - An Introduction

Non-Separable Case III

Solution:
n
X
w= α i yi x i
i=1
à !
X
f (x) = sgn α i yi x >
i x+b
i

KKT conditions:
n
X
0= α i yi
i=1
¡ >
¢
0 = αi (yi w xi + b) − 1 + ξi
0 = (C − αi )ξi

Support vectors: characterized by αi > 0

June 2002 22/44

Support Vector Machines - An Introduction

Non-Separable Case IV
Two types of support vectors:
Recall
¡ >
¢
αi (yi w xi + b) − 1 + ξi = 0
(C − αi )ξi = 0

Margin vectors:

1
0 < αi < C⇒ξi = 0 ⇒ d(xi , H(w, b)) =
kwk

Non-margin vectors: αi = C
? Errors: ξi > 1 Misclassified
? Non-errors: 0 ≤ ξi ≤ 1 Correcty classified
Within margin

June 2002 23/44

Support Vector Machines - An Introduction

Non-Separable Case V

1
3

1
3
2 3

Support Vectors:

1 margin s.v. ξi = 0 Correct

2 non-margin s.v. ξi < 1 Correct (in margin)
3 non-margin s.v. ξi > 1 Error

Problem: Lose clear geometric intuition and

sparsity

June 2002 24/44

Support Vector Machines - An Introduction

Non-linear SVM I

Linear Separability: More likely in high

dimensions

Mapping: Map input into high-dimensional

feature space Φ

Classifier: Construct linear classifier in Φ

Motivation: Appropriate choice of Φ leads to

linear separability.

Non-linearity and high dimension are

essential (Cover ’65)!

Φ : Rd 7→ RD (D À d)
x 7→ Φ(x)

Hyper-plane condition: w> Φ(x) + b = 0

Inner products: x> x 7→ Φ> (x)Φ(x)

June 2002 25/44

Support Vector Machines - An Introduction

Non-Linear SVM II

Decisions in Input and Feature Space:

Problems becomes linearly separable in feature
space (Fig. from Schölkopf & Smola 2002)

R2 x1 x2 input space
✕
Φ: R2 R3
✕ ❍
❍ x12 x22 2 x1x2 feature space
✕ w2 w
✕ w1 3

f(x)
2 2
f (x)=sgn (w1x1+w2x2+w3 2 x1x2+b)
Φ

R3 R2
✕
❍
❍ ✕ ❍
❍
✕
✕
✕ ✕ ✕
✕

Recall for linear SVM

d
X n
X
f (x) = wi x i + b ; w= α i yi x i
i=1 i=1

June 2002 26/44

Support Vector Machines - An Introduction

Non-Linear SVM III

Obtained
n
X
f (x) = α i yi x >
i x+b
i=1

In feature space
n
X
f (x) = αi yi Φ(xi )> Φ(x) + b
i=1

Kernel: A symmetric function K : Rd × Rd 7→ R

Inner product kernels: In addition

K(x, z) = Φ(x)> Φ(z)

Motivation: Φ ∈ RD , where D may be very

large - inner products expensive

June 2002 27/44

Support Vector Machines - An Introduction

Non-Linear Support Vectors IV

Examples:

Linear mapping: Φ(x) = Ax

K(x, z) = (Ax)> (Az) = x> A> Az = x> Bz

Quadratic map: Φ(x) = {(xi xj )}d,d

i,j=1

K(x, z) = (x> z)2

à d !2
X
= x i zi
i=1
d,d
X
= (xi xj )(zi zj )
i,j=(1,1)

Objective: Work directly with kernels,

avoiding mapping Φ

Question: Under what conditions is

K(x, z) = Φ(x)> Φ(z)?

June 2002 28/44

Support Vector Machines - An Introduction

Mercer Kernels I
Assumptions:

1. K(x, z) a continuous symmetric function

2. K is positive definite: for any f ∈ L2 not

identically zero
Z
f (x)K(x, z)f (z)dxdz > 0

Mercer’s Theorem:
∞
X
K(x, z) = λj ψj (x)ψj (z)
j=1
Z
K(x, z)ψj (z)dz = λj ψj (x)

p
Conclusion: Let φj (x) = λj ψj (x), then

K(x, z) = Φ(x)> Φ(z)

June 2002 29/44

Support Vector Machines - An Introduction

Mercer Kernels II
Classifier:
à n
!
X
f (x) = sgn αi yi Φ(xi )> Φ(x) + b
i=1
à n
!
X
= sgn αi yi K(xi , x) + b
i=1

The gain: Implement infinite-dimensional

mapping, but do all calculations in finite
dimension
n
X 1X
maximize αi − αi αj yi yj K(xi , xj )
i=1
2 i,j
n
X
subject to α i yi = 0 ; 0 ≤ αi ≤ C
i=1

Observe: Only difference from linear case is in

the kernel
Optimization task is unchanged!

June 2002 30/44

Support Vector Machines - An Introduction

Kernel Selection I
Simpler Mercer conditions: for any finite set
of points Kij = K(xi , xj ) is positive-definite

v> Kv > 0

Classifier:
à n
!
X
f (x) = sgn αi yi K(xi , x) + b
i=1

Constructing kernels: Assume K1 and K2

kernels, f real-valued function

1. K1 (x, z) + K2 (x, z)

2. K1 (x, z)K2 (x, z)

3. f (x)f (z)

4. K3 (Φ(x), Φ(z))

June 2002 31/44

Support Vector Machines - An Introduction

Kernel Selection II
Explicit construction:

1. p(K1 (x, y)) - polynomial with positive

coefficients

2. exp[K(x, z)]

3. exp(−kx − zk2 /2σ 2 )

Some standard kernels:

Polynomial (x> x + 1)p
¡ 0 2 2
¢
Gaussian exp −kx − x k /2σ
Splines piece-wise polynomial between knots

June 2002 32/44

Support Vector Machines - An Introduction

Kernel Selection III

Adaptive kernels?

? Determine kernel parameters (e.g. width)

using cross-validation

? Improved bounds - take into account

properties of kernels

? Use invariances to constrain kernels

Applications: State of the art in many domains.

Can deal with huge data sets

Hand-writing recognition

Text classification

Bioinformatics

Many more

June 2002 33/44

Support Vector Machines - An Introduction

The Kernel Trick - Summary

? Can be thought of as non-linear similarity

measure

? Can be used with any algorithm that uses

only inner products

? Allows constructing non-linear classifiers

using only linear algorithms

? Can be applied to vectorial and non-vectorial

data (e.g. tree and string structures)

June 2002 34/44

Support Vector Machines - An Introduction

SVM and Penalization I

Recall the linear problem

n
1 2
X
minimize LP (w, ξ) = kwk + C ξi
w,b,ξ 2 i=1

subject to yi (w> xi + b) ≥ 1 − ξi
ξi ≥ 0

Let [u]+ = max(0, u)

One can show equivalence to
( n
)
X
minimize [1 − yi f (xi )]+ + λkwk2
w,b
i=1

subject to f (x) = x> w + b

This has the classic form of Regularization

Theory

empirical error + regularization term

June 2002 35/44

Support Vector Machines - An Introduction

SVM and Penalization II

( n
)
X
minimize [1 − yi f (x)]+ + λkwk2
w,b
i=1

? Maximize distance of points from the hyper-plane

? Correctly classified points are also penalized

[1−yf(x)]
+

I[yf(x) <0]

1 yf(x)

Suggestion: Consider a general formulation

( n )
X
minimize `(yi , f (xi )) + λkwk2
w,b
i=1

subject to f (x) = Φ(x)> w + b

June 2002 36/44

Support Vector Machines - An Introduction

SVM and Penalization III

Recall
( n
)
X
minimize `(yi , f (xi )) + λkwk2 (∗)
w,b
i=1

subject to f (x) = Φ(x)> w + b

Representer Theorem: (special case) For

any loss function `(y, f (x)), the solution of
(∗) is of the form
n
X
f (x) = αi K(xi , x) + b
i=1

Compression bounds: Sparsity improves

generalization! (A version of Occam’s ra-
zor)

June 2002 37/44

Support Vector Machines - An Introduction

SVM Regression I

Data: (x1 , y1 ), . . . , (xn , yn ), yi ∈ R

Loss: Attempt to achieve sparsity

loss

−ε +ε
y−(wx−b)

Achieve: Only badly predicted points con-

tribute

y ξ ε y ξ
ε

ξ ξ

x x

June 2002 38/44

Support Vector Machines - An Introduction

SVM Regression II

Dual problem: Very similar to the case of

classification

Solution: Solve a quadratic optimization

problem in 2n variables αi , αi∗ , i = 1, 2, . . . , n.

n
X
f (x) = (αi∗ − αi )K(xi , x) + b
i=1

Sparsity: Only data for which αi 6= αi∗

contribute. This occurs only if

|f (xi ) − yi | ≥ ² (outside tube)

June 2002 39/44

Support Vector Machines - An Introduction

SVM Regression III

Effect of ²:

² = 0.1, 0.2, 0.5

Improved regression: ²-adaptive

Left - zero noise, right - σ = 0.2

1 1

0 0

0 1 0 1

Figures: from Schölkopf and Smola 2002

June 2002 40/44

Support Vector Machines - An Introduction

Practically Useful Bounds?

Practically useful bound can be obtained using:

? Data-dependent complexities

? Specific learning algorithms

? A great deal of ingenuity

40
Test error
35 Span prediction

30

25
Error

20

15

10

0
−6 −4 −2 0 2 4 6
Log sigma


   
 
"!

#
$%
&

36
Test error
34 Span prediction

32

30
Error

28

26

24

22

20
−2 0 2 4 6 8 10 12
Log C
'$(
)* ,+- .$./0
&120

3/4!


$%
&

From: “Bounds on Error Expectations for Support

Vector Machines”, V. Vapnik and O. Chapelle, 2001
(200 examples in (b))

June 2002 41/44

Support Vector Machines - An Introduction

Summary I
Advantages

? Systematic implementation through quadratic

programming (very efficient implementations
exist)

? Excellent data-dependent generalization

bounds exist

? Regularization built into cost function

? Statistical performance independent of dim.

of feature space

? Theoretically related to widely studied fields

of regularization theory and sparse
approximation

? Fully adaptive procedures available for

determining hyper-parameters

June 2002 42/44

Support Vector Machines - An Introduction

Summary II
Drawbacks

? Treatment of non-separable case somewhat

heuristic

? Number of support vectors may depend

strongly on the kernel type and and the
hyper-parameters

? Systematic choice of kernels is difficult (prior

information) - some ideas exist

? Optimization may require clever heuristics for

large problems

June 2002 43/44

Support Vector Machines - An Introduction

Summary III
Extensions

? Online algorithms

? Systematic choice of kernels using generative

statistical models

? Applications to

– Clustering
– Non-linear principal component analysis
– Independent component analysis

? Generalization bounds constantly improving

(some even practically useful!)

June 2002 44/44