31
1 Optimization Problem Based on L 2,1 - norms Xiaohong Chen 19-10-201 2

Optimization Problem Based on L 2,1 -norms

Embed Size (px)

DESCRIPTION

Optimization Problem Based on L 2,1 -norms. Xiaohong Chen 19-10-2012. Outline. Efficient and robust feature selection via joint l 2,1 -norm minimzation Robust and discriminative distance for multi-instance learning Its application…. Outline. - PowerPoint PPT Presentation

Citation preview

Page 1: Optimization Problem Based on L 2,1 -norms

1

Optimization Problem Based on L2,1-norms

Xiaohong Chen 19-10-2012

Page 2: Optimization Problem Based on L 2,1 -norms

2

Outline Efficient and robust feature selection via

joint l2,1-norm minimzation Robust and discriminative distance for m

ulti-instance learning Its application…

Page 3: Optimization Problem Based on L 2,1 -norms

3

Outline Efficient and robust feature selection via

joint l2,1-norm minimization Robust and discriminative distance for

multi-instance learning Its application…

Page 4: Optimization Problem Based on L 2,1 -norms

4

Efficient and robust feature selection via joint l2,1-norm minimzation

Page 5: Optimization Problem Based on L 2,1 -norms

5

Robust Feature Selection Based on l21-norm

Given training data {x1, x2,…, xn} and the associated class labels {y1,y2,…, yn}

Least square regression solves the following optimizaiton problem to obtain the projection matrix W

Add a regularization R(W) to the robust version of LS,

Page 6: Optimization Problem Based on L 2,1 -norms

6

Robust Feature Selection Based on l21-norm

Possible regularizations

Ridge regularization

Lasso regularization

Lasso regularization

Penalize all c regression coefficients corresponding to a single feature as a whole

Page 7: Optimization Problem Based on L 2,1 -norms

7

Robust Feature Selection Based on l21-norm

Page 8: Optimization Problem Based on L 2,1 -norms

8

Robust Feature Selection Based on l21-norm

Denote

(14)

Page 9: Optimization Problem Based on L 2,1 -norms

9

Robust Feature Selection Based on l21-norm

Then we have

(19)

Page 10: Optimization Problem Based on L 2,1 -norms

10

The iterative algorithm to solve problem (14)

Theorem1: The algorithm will monotonically decrease the objective of the problem in Eq.(14) in each iteration, and converge to the globaloptimum of the problem.

Page 11: Optimization Problem Based on L 2,1 -norms

11

Proof of theorem12 2

2 2

a ba

b b 2 22ab a b

2 2

2 2

a ba b

b b

u u

Page 12: Optimization Problem Based on L 2,1 -norms

12

Proof of theorem1

Page 13: Optimization Problem Based on L 2,1 -norms

13

(1)

(2)

(1)+(2)

Page 14: Optimization Problem Based on L 2,1 -norms

14

Outline Efficient and robust feature selection via

joint l2,1-norm minimization Robust and discriminative distance for

multi-instance learning Its application…

Page 15: Optimization Problem Based on L 2,1 -norms

15

Robust and discriminative distance for multi-instance learning

Page 16: Optimization Problem Based on L 2,1 -norms

16

Multi-instance learning

多示例学习中,训练集由若干个具有概念标记的包 (bag) 组成,每个包包含若干个没有概念标记的示例。若一个包中至少有一个正例,则该包被标记为正 (positive) ,若一个包中所以示例都是反例,则该包被标记为反 (negative), 通过对训练包的学习,希望学习系统尽可能正确地对训练集之外的包的概念标记进行预测。

Page 17: Optimization Problem Based on L 2,1 -norms

17

The illustration of MIL

Page 18: Optimization Problem Based on L 2,1 -norms

18

Notations

Given N training bags and K conceptual classes.

Each bag contains a number of instances

Given the class memberships of the input data, denoted as

Page 19: Optimization Problem Based on L 2,1 -norms

19

Notations

First, we represent every class as a super-bag that comprises the

instances of all its training , where

Page 20: Optimization Problem Based on L 2,1 -norms

20

Objective to learn class specific distance metrics

For a given class, Ck,, we solve the following optimization problem:

Page 21: Optimization Problem Based on L 2,1 -norms

21

Algorithm and its analysis

Page 22: Optimization Problem Based on L 2,1 -norms

22

Algorithm and its analysis

Page 23: Optimization Problem Based on L 2,1 -norms

23

Algorithm and its analysis

Page 24: Optimization Problem Based on L 2,1 -norms

24

Algorithm and its analysis

On the other hand,

Page 25: Optimization Problem Based on L 2,1 -norms

25

Algorithm and its analysis

Page 26: Optimization Problem Based on L 2,1 -norms

26

Algorithm and its analysis

Page 27: Optimization Problem Based on L 2,1 -norms

27

Algorithm and its analysis

Therefore, the objective value of the problem of (6) is decreased in each iteration till convergences.

Page 28: Optimization Problem Based on L 2,1 -norms

28

Outline Efficient and robust feature selection via

joint l2,1-norm minimzation Robust and discriminative distance for m

ulti-instance learning Its application…

Page 29: Optimization Problem Based on L 2,1 -norms

29

Its application

, 2,1

2,1,

( )( )

min min( )( )

i j

i j

T Ti j i j

x x

T T Wi j i j

x x

W x x x x WAW

W x x x x W BW

同类

不同类

For example:

2,1

2,1 2,1

minW

AW

BW CW

Page 30: Optimization Problem Based on L 2,1 -norms

30

[1]F.Nie, D.Xu, X.Cai, and C.Ding. Efficient and robust feature selection via

joint l2,1-norm minimzation. NIPS 2010.

[2] H.Wang, F.Nie and H.Huang. Robust and discriminative distance for multi-

instance learning, CVPR 2012: 2919-2924

Reference

Page 31: Optimization Problem Based on L 2,1 -norms

31

Thanks! Q&A