least squares support vector machine classifiers j.a.k. suykens and j. vandewalle presenter: keira...

15
Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Upload: andrea-lynch

Post on 03-Jan-2016

224 views

Category:

Documents


0 download

TRANSCRIPT

Page 1: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Least Squares Support Vector Machine ClassifiersJ.A.K. Suykens and J. Vandewalle

Presenter: Keira (Qi) Zhou

Page 2: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Outline• Background• Classic Support Vector Machine (SVM)• Optimization for SVM• Linear Programming vs. Quadratic Programming

• Least Square Support Vector Machine (LS-SVM)• Optimization for LS-SVM

• Comparison

2

Page 3: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

3

Page 4: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

L1: wx + b = 1

L2: wx + b = -1

wx + b = 0

Support Vector Machine

Margin:2/|w|Maximize Margin => Minimize |w|

4

Support Vectors

Save this in your memory buffer for now

Page 5: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Support Vector Machine (Cont’d)• What if…

5

Page 6: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

• Introduce slack variables• Allow some mistakes

6

Support Vector Machine (Cont’d)

Page 7: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Optimization for SVM• Formulation

• Lagrange Multiplier

• Take the derivatives and optimality condition

7

Page 8: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Optimization for SVM (Cont’d)• End up solving a quadratic programming problem

• We first find α, then use α to calculate w and b

8

Page 9: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Linear Programming vs. Quadratic Programming

• Linear Programming• Linear objective function• Linear constraints

• Quadratic Programming• Quadratic objective function• Linear constraints

9

Page 10: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

How much one may simplify the SVM formulation without losing any of its

advantages?

10

SO…

Page 11: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Least Square Support Vector Machine

11

Similar to

regression?

Page 12: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Optimization for LS-SVM• Lagrange Multiplier

12

Page 13: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Optimization for LS-SVM (Cont’d)• Now take the derivative together with optimality condition,

we end up with a set of linear equations instead of quadratic programming

#EasyToSolve ! 13

Page 14: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Comparison• How much one may simplify the SVM formulation without

losing any of its advantages?

• Experiments on 3 dataset [1]

ALL LEUKEMIA ALLAML3

SVM 96.98 97.69 95.97

LS-SVM 97.33 97.00 93.83

[1] Ye, Jieping, and Tao Xiong. "SVM versus least squares SVM." International Conference on Artificial Intelligence and Statistics. 2007.

14

Page 15: Least Squares Support Vector Machine Classifiers J.A.K. Suykens and J. Vandewalle Presenter: Keira (Qi) Zhou

Question?

15