An Introduction to Support Vector Machineszhou/568/SVM.pdf · 2016. 3. 4. · Support Vector Machine (SVM) ! A classifier derived from statistical learning theory by Vapnik, et al.

An Introduction to Support Vector Machine

Support Vector Machine (SVM)

n  A classifier derived from statistical learning theory by Vapnik, et al. in 1992

n  SVM became famous when, using images as input, it gave accuracy comparable to neural-network with hand-designed features in a handwriting recognition task

n  Currently, SVM is widely used in object detection & recognition, content-based image retrieval, text recognition, biometrics, speech recognition, etc.

n  Also used for regression

Outline

n  Linear Discriminant Function n  Large Margin Linear Classifier n  Nonlinear SVM: The Kernel Trick

Linear Discriminant Function

n  g(x) is a linear function:

( ) Tg b= +x w x

x1

x2

wT x + b < 0

wT x + b > 0

n  A hyper-plane in the feature space

n  (Unit-length) normal vector of the hyper-plane:

=wnw

n wT x + b = 0

n  How would you classify these points using a linear discriminant function in order to minimize the error rate?

Linear Discriminant Function denotes +1

denotes -1

x1

x2

n  Infinite number of answers!



denotes -1

x1

x2




denotes -1

x1

x2


x1

x2 n  How would you classify these points using a linear discriminant function in order to minimize the error rate?


denotes -1


n  Which one is the best?

Large Margin Linear Classifier

“safe zone” n  The linear discriminant

function (classifier) with the maximum margin is the best

n  Margin is defined as the width that the boundary could be increased by before hitting a data point

n  Why it is the best? q  Robust to outliners and thus

strong generalization ability q  Good according to PAC

(Probably Approximately Correct) theory.

Margin

x1

x2

denotes +1

denotes -1

Maximum Margin Classification n  Distance from point xi to the

hyperplane is: n  Examples closest to the

hyperplane are support vectors.

n  Margin M of the classifier is the distance between support vectors on both sides.

n 

wxw br i

T +=

x1

x2 Margin

x+

x+

x- n

Support Vectors

Only support vectors matter; other training points are ignorable.

M


n  Given a set of data points:

n  With a scale transformation on both w and b, the above is equivalent to

x1

x2

denotes +1

denotes -1

{( , )}, 1,2, ,i iy i n=x L , where

For 1, 1

For 1, 1

Ti i

Ti i

y by b= + + ≥

= − + ≤ −

w xw x

M

wTxi + b ≥ M/2 if yi = +1

wTxi + b ≤ - M/2 if yi = -1


n  We know that

n  The margin width is:

x1

x2

denotes +1

denotes -1

1 1

T

T

bb

+

−

+ =

+ = −

w xw x

Margin

x+

x+

x-

( )2 ( )

M + −

+ −

= − ⋅

= − ⋅ =

x x nwx xw w

n

Support Vectors

M

Thus wT (x+-x-) = 2


n  Formulation:

x1

x2

denotes +1

denotes -1

Margin

x+

x+

x- n

such that

2maximize w

For 1, 1

For 1, 1

Ti i

Ti i

y by b= + + ≥

= − + ≤ −

w xw x


n  Formulation:

x1

x2

denotes +1

denotes -1

Margin

x+

x+

x- n

21minimize 2w

such that

For 1, 1

For 1, 1

Ti i

Ti i

y by b= + + ≥

= − + ≤ −

w xw x


n  Formulation:

x1

x2

denotes +1

denotes -1

Margin

x+

x+

x- n ( ) 1T

i iy b+ ≥w x

21minimize 2w

such that

Solving the Optimization Problem

( ) 1Ti iy b+ ≥w x

21minimize 2w

s.t.

Quadratic programming

with linear constraints

( )2

1

1minimize ( , , ) ( ) 12

nT

p i i i ii

L b y bα α=

= − + −∑w w w x

s.t.

Lagrangian Function

0iα ≥


( )2

1

1minimize ( , , ) ( ) 12

nT

p i i i ii

L b y bα α=

= − + −∑w w w x

s.t. 0iα ≥

0pLb

∂=

∂

0pL∂ =∂w 1

n

i i ii

yα=

=∑w x

10

n

i ii

yα=

=∑


( )2

1

1minimize ( , , ) ( ) 12

nT

p i i i ii

L b y bα α=

= − + −∑w w w x

s.t. 0iα ≥

1 1 1

1maximize 2

n n nT

i i j i j i ji i j

y yα αα= = =

−∑ ∑∑ x x

s.t. 0iα ≥1

0n

i ii

yα=

=∑, and

Lagrangian Dual Problem


n  The solution has the form:

( )( ) 1 0Ti i iy bα + − =w x

n  From KKT (Karush–Kuhn–Tucker) condition, we know:

n  Thus, only support vectors have 0iα ≠

1 SV

n

i i i i i ii i

y yα α= ∈

= =∑ ∑w x x

get b from yk (wT xk +b)−1= 0,

where xk is any support vector

x1

x2

x+

x+

x-

Support Vectors

Thus, b = yi - Σαiyixi Txk for any αk > 0


g (x) = wT x +b = αi yi xiT x

i∈SV∑ +b

n  The linear discriminant function is:

n  Notice it relies on a dot product between the test point x and the support vectors xi

n  Also keep in mind that solving the optimization problem involved computing the dot products xi

Txj between all pairs of training points

That is, no need to compute w explicitly for classification.


n  What if data is not linear separable? (noisy data, outliers, etc.)

n  Slack variables ξi can be added to allow mis-classification of difficult or noisy data points

x1

x2

denotes +1

denotes -1

1ξ2ξ


n  Formulation:

( ) 1Ti i iy b ξ+ ≥ −w x

2

1

1minimize 2

n

ii

C ξ=

+ ∑w

such that 0iξ ≥

n  Parameter C can be viewed as a way to control over-fitting: it “trades off” the relative importance of maximizing the margin and fitting the training data. n  For large values of C, the optimization will choose a smaller-margin

hyperplane if that hyperplane does a better job of getting all the training points classified correctly. Conversely, a very small value of C will cause the optimizer to look for a larger-margin separating hyperplane, even if that hyperplane misclassifies more points.


n  Formulation: (Lagrangian Dual Problem)

1 1 1

1maximize 2

n n nT

i i j i j i ji i j

y yα αα= = =

−∑ ∑∑ x x

such that

0 i Cα≤ ≤

10

n

i ii

yα=

=∑


n  Again, xi with non-zero αi will be support vectors.

n  Solution to the dual problem is:

b= yk(1- ξk) - Σαiyixi

Txk for any k s.t. αk>0

Again, we don’t need to compute w explicitly for classification:

1 SV

n

i i i i i ii i

y yα α= ∈

= =∑ ∑w x x

g (x) = wT x +b = αi yi xiT x

i∈SV∑ +b

Non-linear SVMs n  Datasets that are linearly separable with noise work out great:

0 x

0 x

x2

0 x

n  But what are we going to do if the dataset is just too hard?

n  How about… mapping data to a higher-dimensional space:

Non-linear SVMs: Feature Space n  General idea: the original input space can be mapped to

some higher-dimensional feature space where the training set is separable:

Φ: x → φ(x)

Nonlinear SVMs: The Kernel Trick n  With this mapping, our discriminant function is now:

SV( ) ( ) ( ) ( )T T

i ii

g b bφ αφ φ∈

= + = +∑x w x x x

n  No need to know this mapping explicitly, because we only use the dot product of feature vectors in both the training and test.

n  A kernel function is defined as a function that corresponds to a dot product of two feature vectors in some expanded feature space:

( , ) ( ) ( )Ti j i jK φ φ≡x x x x

Nonlinear SVMs: The Kernel Trick

2-dimensional vectors x=[x1 x2]; let K(xi,xj)=(1 + xi

Txj)2,

Need to show that K(xi,xj) = φ(xi)

Tφ(xj):

K(xi,xj)=(1 + xiTxj)2

,

= 1+ xi12xj1

2 + 2 xi1xj1 xi2xj2+ xi2

2xj22 + 2xi1xj1 + 2xi2xj2

= [1 xi12 √2 xi1xi2 xi2

2 √2xi1 √2xi2]T [1 xj12 √2 xj1xj2 xj2

2 √2xj1 √2xj2] = φ(xi)

Tφ(xj), where φ(x) = [1 x12 √2 x1x2 x2

2 √2x1 √2x2]

n  An example:

This slide is courtesy of www.iro.umontreal.ca/~pift6080/documents/papers/svm_tutorial.ppt

Nonlinear SVMs: The Kernel Trick

q  Linear kernel:

2

2( , ) exp( )2i j

i jKσ

−= −

x xx x

( , ) Ti j i jK =x x x x

( , ) (1 )T pi j i jK = +x x x x

0 1( , ) tanh( )Ti j i jK β β= +x x x x

n  Examples of commonly-used kernel functions:

q  Polynomial kernel:

q  Gaussian (Radial-Basis Function (RBF) ) kernel:

q  Sigmoid:

Mercer’s theorem: Every semi-positive definite symmetric function is a kernel.

Nonlinear SVM: Optimization n  Formulation: (Lagrangian Dual Problem)

1 1 1

1maximize ( , )2

n n n

i i j i j i ji i j

y y Kα αα= = =

−∑ ∑∑ x x

such that 0 i Cα≤ ≤

10

n

i ii

yα=

=∑

n  The solution of the discriminant function is

g (x) = αi yi K (xi ,x)i∈SV∑ +b

n  The optimization technique is the same.

Support Vector Machine: Algorithm

n  1. Choose a kernel function

n  2. Choose a value for C

n  3. Solve the quadratic programming problem (many software packages available)

n  4. Construct the discriminant function from the support vectors

Some Issues n  Choice of kernel - Gaussian or polynomial kernel is default - if ineffective, more elaborate kernels are needed - domain experts can give assistance in formulating appropriate

similarity measures n  Choice of kernel parameters - e.g. σ in Gaussian kernel - σ is the distance between closest points with different classifications - In the absence of reliable criteria, applications rely on the use of a

validation set or cross-validation to set such parameters. n  Optimization criterion – Hard margin v.s. Soft margin - a lengthy series of experiments in which various parameters are

tested

This slide is courtesy of www.iro.umontreal.ca/~pift6080/documents/papers/svm_tutorial.ppt

Summary: Support Vector Machine

n  1. Large Margin Classifier q  Better generalization ability & less over-fitting

n  2. The Kernel Trick q  Map data points to higher dimensional space in

order to make them linearly separable. q  Since only dot product is used, we do not need to

represent the mapping explicitly.

Demo of LibSVM

http://www.csie.ntu.edu.tw/~cjlin/libsvm/ �

References on SVM and Stock Prediction

http://www.svms.org/finance/HuangNakamoriWang2005.pdf http://cs229.stanford.edu/proj2012/ShenJiangZhang-StockMarketForecastingusingMachineLearningAlgorithms.pdf http://research.ijcaonline.org/volume41/number3/pxc3877555.pdf and other references online …

An Introduction to Support Vector Machineszhou/568/SVM.pdf · 2016. 3. 4. · Support Vector Machine (SVM) ! A classifier derived from statistical learning theory by Vapnik, et al.

Documents