Support Vector Machines

Xiaojin Zhu
[email protected]
Computer Sciences Department
University of Wisconsin, Madison

[ Based on slides from Andrew Moore http://www.cs.cmu.edu/~awm/tutorials]

slide 1

The no-math verison

Support vector machines

slide 2

Courtesy of Steve Chien of NASA/JPL

Lake Mendota, Madison, WI

Lake Mendota, Wisconsin

Identify areas of land cover (land, ice,

water, snow) in a scene
Three algorithms:

Scientist manually derived

Automatic best ratio

Support Vector Machine (SVM)

Classifier

Expert
Derived

Automated
Ratio

SVM

cloud

45.7%

43.7%

58.5%

ice

60.1%

34.3%

80.4%

land

93.6%

94.7%

94.0%

snow

63.5%

90.4%

71.6%

water

84.2%

74.3%

89.1%

unclassified

45.7%

Visible
Image

Expert
Labeled

Expert
Derived

Automated
Ratio

slide 3

SVM

Support vector machines

The state-of-the-art classifier

Class labels
denotes +1

denotes -1

How would you

classify this data?

slide 4

Linear classifier

If all you can do is to draw a straight line, the linear

decision boundary

slide 5

Linear classifier

Another OK decision boundary

slide 6

slide 7

Any of these
would be fine..
..but which is the
best?

slide 8

The margin

Margin: the width that the boundary can be increased

before hitting a data point

slide 9

SVM: maximize margin

The simplest SVM (linear SVM) is the linear classifier

with the maximum margin

Linear SVM
slide 10

SVM: linearly non-separable data

What if the data is not linearly separable?

slide 11

SVM: linearly non-separable data

Two solutions:
Allow a few points on the wrong side (slack
variables), and/or
Map data to a higher dimensional space, do linear
classification there (kernel)

slide 12

SVM: more than two classes

N class problem: Split the task into N binary tasks:

Class 1 vs. the rest (class 2N)
Class 2 vs. the rest (class 1, 3N)

Class N vs. the rest
Finally, pick the class that put the point furthest into
the positive region.

slide 13

SVM: get your hands on it

There are many implementations

http://www.support-vector.net/software.html
http://svmlight.joachims.org/

You dont have to know the rest of the class in order

to use SVM.

end of class
(But you want to know more, dont you?)

slide 14

The math verison

Support vector machines

slide 15

Vector

A vector W in d-dimensional space is a list of d

numbers, e.g. W=(-1,2)
A vector is a vertical column. is matrix transpose
Vector is a line segment, with arrow, in space
The norm of a vector ||W|| is its length
||W|| = sqrt(W W)
W
Inner product: XY = XiYi

What d-dimensional point X

makes W X = 0?

slide 16

Lines

W X is a scalar (single number): Xs projection onto W

WX=0 specifies the set of points X, which is the line
perpendicular to W, and intersects at (0,0)

WX=1, or WX-1=0

WX=0
WX=-1, or WX+1 = 0

1/||W||

WX=1 is the line parallel to WX=0, shifted by 1/||W||

What if the boundary doesnt go through origin?
slide 17

SVM boundary and margin

Want: find W, b (offset) such that

all positive training points (X,Y=1) in the red zone,
all negative ones (X, Y=-1) in the blue zone,
the margin M is maximized
Plus-Plane
Classifier Boundary
Minus-Plane

2/||W||

M=2/||W|| (why?)
How do we find such W,b?
slide 18

SVM as constrained optimization

Variables: W, b
Objective function: maximize the margin M=2/||W||
Equiv. to minimize ||W||, or ||W||2=WW, or WW

Assume N training points (Xi, Yi), Yi = 1 or -1

Subject to each training point on the correct side (the
constraint). How many constraints do we have?
Plus-Plane
Classifier Boundary
Minus-Plane

2/||w||

slide 19

SVM as constrained optimization

Variables: W, b
Objective function: maximize the margin M=2/||W||
Equiv. to minimize ||W||, or ||W||2=WW, or WW

Assume N data points (Xi, Yi), Yi = 1 or -1

Subject to each training point on the correct side (the
constraint). How many constraints do we have? N
WXi + b >= 1, if Yi=1
WXi + b <= -1, if Yi=-1
we can unify them: Yi(WXi+b) >= 1
We have a continuous constrained optimization
problem! What do we do?
slide 20

SVM as QP

minW,b WW
Subject to Yi (WXi + b) >= 1, for all i
Objective is convex, quadratic
Linear constraints
This problem is known as Quadratic Program (QP),
for which efficient global solution algorithms exist.
Plus-Plane
Classifier Boundary
Minus-Plane

2/||W||

slide 21

Non-separable case

What about this?

Can we insist on Yi (WXi + b) >= 1, for all i?

slide 22

Trick #1: slack variables

Relax the constraints allow a few bad apples

For a given linear boundary W, b, we can compute
how far off into the wrong side a bad point is

e2

e11

e7

Heres how we relax the constraints:

Yi (WXi + b) >= 1- ei
slide 23

Trick #1: SVM with slack variables

minW,b,e WW + C i ei
Subject to

Trade-off parameter

Yi (WXi + b) >= 1- ei for all i

ei>=0 for all i (why?)

e2

e11

e7

slide 24

Trick #1: SVM with slack variables

minW,b,e WW + C i ei
Subject to

Trade-off parameter

Yi (WXi + b) >= 1- ei for all i

ei>=0 for all i

Originally we optimize variables 11

e2
W (d-dimensional vector),
b
Now we optimize W, b, e1 eN

e7

Now we have 2N constraints

Still a QP (soft-margin SVM)
slide 25

Another look at non-separable case

Heres another non-separable dataset, in 1dimensional space

We can of course use slack variables but heres
another trick!

x=0

slide 26

Trick #2:Map data to high dimensional space

We can map data x from 1-D to 2-D by x (x, x2)

x=0
slide 27

Trick #2:Map data to high dimensional space

We can map data x from 1-D to 2-D by

x (x)=(x, x2)
Now the data is
linearly separable in
the new space!
We can run SVM in
the new space
The linear boundary
in the new space
corresponds to a nonlinear boundary in the
old space
x=0
slide 28

Another example

( x1 , x2 ) ( x1 , x2 , x x )
2
1

2
2

slide 29

Trick #2:Map data to high dimensional space

In general we might want to map an already highdimensional X=(x1, x2, , xd) into some much higher,
even infinite dimensional space (x)
Problems:
How do you represent infinite dimensions?
We need to learn (amount other things) W, which
lives in the new space learning a large (or
infinite) number of variables in QP is not a good
idea.

x=0
slide 30

Trick #2: kernels

We will do several things:

Convert it into a equivalent QP problem, which
does not use W, or even (X) alone!
It only uses the inner product (X) (Y), where X and Y
are two training points. The solution also only uses such
inner product
It still seems infeasible to compute for high (infinite)
dimensions

But there are smart ways to compute such inner

product, known as kernel (a function on two var.)

kernel(X,Y) inner product (X) (Y)

Why should you care:

One kernel, one new (higher dimensional) space
You will impress friends at cocktail parties
slide 31

Prepare to bite the bullet

Heres the original QP formula

minW,b WW
Subject to Yi (WXi + b) >= 1, for all i (N constraints)
Remember Lagrange multipliers?
We have them

L = WW ai [Yi (WXi + b) 1]

here because
those are
inequality
constraints

Subject to ai >= 0, for all i

We want the gradient of L to vanish with respect to W,
b, a. Try this and youll get
W = ai Yi Xi

ai Yi = 0
Put them back into the Lagrangian L
slide 32

Biting the bullet

max{ai} ai i,jai aj Yi Yj Xi Xj
Subject to

ai >= 0, for all I

ai Yi = 0
This is an equivalent QP problem (the dual)
Before we optimize W (d variables), now we optimize
a (N variables): which is better?
X only appears in the inner product

slide 33

Biting the bullet

max{ai} ai i,jai aj Yi Yj (Xi) (Xj)
Subject to

ai >= 0, for all I

ai Yi = 0

If we map X to
new space
(X)

This is an equivalent QP problem

Before we optimize W (d variables), now we optimize
a (N variables): which is better?
X only appears in the inner product

slide 34

Biting the bullet

max{ai} ai i,jai aj Yi Yj K(Xi, Xj)
Subject to

ai >= 0, for all I

ai Yi = 0

If we map X to
new space
(X)

This is an equivalent QP problem

Before we optimize W (d variables), now we optimize
a (N variables): which is better?
X only appears in the inner product

Kernel K(Xi,Xj) = (Xi)

(Xj)

slide 35

Whats special about kernel

Say data is two dimensional: s=(s1,s2)

We decide to use a particular mapping into 6
dimensional space
(s)=(s12, s22, 2 s1s2, s1, s2, 1)
Let another point be t=(t1,t2).
(s) (t)=s12 t12 + s22 t22 + 2 s1s2t1t2 + s1t1 + s2t2 + 1

slide 36

Whats special about kernel

Say data is two dimensional: s=(s1,s2)

We decide to use a particular mapping into 6
dimensional space
(s)=(s12, s22, 2 s1s2, 2 s1, 2 s2, 1)
Let another point be t=(t1,t2).
(s) (t)=s12 t12 + s22 t22 + 2 s1s2t1t2 + 2 s1t1 + 2 s2t2 + 1

Let the kernel be K(s,t) = (st+1)2

Verify that they are the same. We saved
computation.

slide 37

Kernels

You ask: Is there such a good K for any I pick?

The inverse question: Given some K, is there a so
that K(X,Y) = (X) (Y)?
Mercers condition: the inverse question is true

This is positive semi-definiteness, if you must know

may be infinite dimensional; we may not be able to
explicitly write down

slide 38

Some frequently used kernels

Linear kernel: K(X,Y) = XY

Quadratic kernel: K(X,Y) = (XY+1)2
Polynomial kernel: K(X,Y) = (XY+1)n
Radial Basis Function kernel: K(X,Y) = exp(- ||X-Y||2/)
Many, many other kernels
Hacking with SVM: create various kernels, hope their
space is meaningful, plug them into SVM, pick the
one with good classification accuracy (equivalent to
feature engineering)
Kernel summary: QP of size N, nonlinear SVM in the
original space, new space possibly high/infinite dim,
efficient if K is easy to compute
Kernel can be combined with slack variables
slide 39

Why the name support vector machines

max{ai} ai i,jai aj Yi Yj K(Xi, Xj)
Subject to

ai >= 0, for all I

ai Yi = 0
The decision boundary is
f(Xnew)=W Xnew + b= ai Yi Xi Xnew + b
In practice, many as will be zero in the solution!
Those few X with a>0 lies on the margin (blue or
red lines), they are the support vectors

slide 40

What you should know

The intuition, where to find software

Vector, line, length
Margin
QP with linear constraints
How to handle non-separable data
Slack variables
Kernels new feature space

Ref: A Tutorial on Support Vector Machines for Pattern

Recognition (1998) Christopher J. C. Burges

slide 41