support vector machine

33
SEN – 935 DATA MINING Anandha L Ranganathan Support Vector Machine 17. Support Vector Machine 1

Upload: anandha-l-ranganathan

Post on 26-Jan-2015

497 views

Category:

Technology


1 download

DESCRIPTION

Support Vector machine

TRANSCRIPT

Page 1: Support Vector machine

SEN – 935 DATA MINING

Anandha L Ranganathan

Support Vector Machine

17. Support Vector Machine 1

Page 2: Support Vector machine

Pre 1980s - Almost all learning methods learned linear decision surfaces. - Linear learning methods have nice theoretical properties

1980s - Almost all learning methods learned linear decision surfaces. - Linear learning methods have nice theoretical properties

1990’s - Efficient learning algorithms for non-linear functions based on

computational learning theory developed - Nice theoretical properties.

History

17. Support Vector Machine 2

Page 3: Support Vector machine

Two independent developments within last decade

– Computational learning theory– New efficient separability of non-linear functions that use “kernel

functions”

The resultant learning algorithm is optimization algorithm rather than a greedy search.

What is greedy search ?

History

17. Support Vector Machine 3

Page 4: Support Vector machine

Find largest sum by traversing through path.

Greedy search

17. Support Vector Machine 4

Page 5: Support Vector machine

Find largest sum by traversing through path.

Greedy search

17. Support Vector Machine 5

Page 6: Support Vector machine

Find largest sum by traversing through path.

Greedy search

17. Support Vector Machine 6

Page 7: Support Vector machine

A system receives data as input. Output a function that can be predict some

features of future data.

Learning Theory

f x yest

f(x,w,b) = sign(w. x - b)

17. Support Vector Machine 7

Page 8: Support Vector machine

Not affected by local minima. Do not suffer from the curse of dimensionality. Have modular design that allows one to

separately implement and design other component.

Various properties of the SVM solution help avoid over fitting, even in very high dimensional feature spaces

Features of SVM's

17. Support Vector Machine 8

Page 9: Support Vector machine

Support vectors are data points that lie closes to the decision surface.

But they are difficult to classify. They have direct bearing of optimum location

on the surface.

Support Vectors

17. Support Vector Machine 9

Page 10: Support Vector machine

d1 and d2 are 2 vectors. And sum of their distance is d1+d2=q.

d1=2x+5y and d2=3x+2y q=d1+d2=5x+7y

Vector Space – Primer

17. Support Vector Machine 10

Page 11: Support Vector machine

Computing the inner products between the vectors in the featured space.

Kernal – Primer

17. Support Vector Machine 11

Page 12: Support Vector machine

Linear Classifiers

denotes +1

denotes -1

How would you classify this data?

17. Support Vector Machine 12

Page 13: Support Vector machine

denotes +1

denotes -1

How would you classify this data?

Linear Classifiers

17. Support Vector Machine 13

Page 14: Support Vector machine

denotes +1

denotes -1

How would you classify this data?

Linear Classifiers

17. Support Vector Machine 14

Page 15: Support Vector machine

denotes +1

denotes -1

How would you classify this data?

Linear Classifiers

17. Support Vector Machine 15

Page 16: Support Vector machine

denotes +1

denotes -1

How would you classify this data?

Linear Classifiers

17. Support Vector Machine 16

Page 17: Support Vector machine

denotes +1

denotes -1

Any of these would be fine..

..but which is best?

Linear Classifiers

17. Support Vector Machine 17

Page 18: Support Vector machine

Classifier Margin

denotes +1

denotes -1 Define the margin of a linear classifier as the width that the boundary could be increased by before hitting a datapoint.

17. Support Vector Machine 18

Page 19: Support Vector machine

Maximum Margin

denotes +1

denotes -1 The maximum margin linear classifier is the linear classifier with the maximum margin.

This is the simplest kind of SVM (Called an LSVM)

Linear SVM17. Support Vector Machine 19

Page 20: Support Vector machine

Formulating SVM

17. Support Vector Machine 20

Page 21: Support Vector machine

Formulating SVM

17. Support Vector Machine 21

Page 22: Support Vector machine

Formulating SVM

17. Support Vector Machine 22

Page 23: Support Vector machine

Idea: map to higher dimensional feature space

Kernal - polynomial

17. Support Vector Machine 23

Page 24: Support Vector machine

License Plate Recognition

17. Support Vector Machine 24

Page 25: Support Vector machine

Pre-process the image of number plate. Segment the image into several parts of

which each contains only a single character.

License Plate Recognition

17. Support Vector Machine 25

Page 26: Support Vector machine

Extract the feature vector of each normalized candidate

Recognizes the single character (a digit or a letter) by the set of SVMs trained in advance.

License Plate Recognition

17. Support Vector Machine 26

Page 27: Support Vector machine

License Plate Recognition

17. Support Vector Machine 27

Page 28: Support Vector machine

License Plate Recognition

17. Support Vector Machine 28

Page 29: Support Vector machine

License Plate Recognition

17. Support Vector Machine 29

Page 30: Support Vector machine

License Plate Recognition

17. Support Vector Machine 30

Page 31: Support Vector machine

License Plate Recognition

If there are no more unclassified samples, then STOP. Otherwise, then repeat the process of recognition of character.

Add these test samples into their corresponding database for further training.

Recognize number plate by bringing all characters used together

17. Support Vector Machine 31

Page 32: Support Vector machine

SVM is widely used as classify spam detection in the market.

It supports for Linear and Non-Linear spectrum.

Conclusion

17. Support Vector Machine 32

Page 33: Support Vector machine

http://www.cs.ucf.edu/courses/cap6412/fall2009/papers/Berwick2003.pdf

http://physiology.med.cornell.edu/people/banfelder/qbio/resources_2011/2011_Leslie.pdf

http://physiology.med.cornell.edu/people/banfelder/qbio/resources_2011/2011_Leslie.pdf

http://www.cs.columbia.edu/~kathy/cs4701/documents/jason_svm_tutorial.pdf

http://www.slideshare.net/wltongxing/svm-12978262

17. Support Vector Machine 33