+ All Categories
Home > Documents > Sparse Screening for Exact Data Reduction · Sparse Screening for Exact Data Reduction ... Sparse...

Sparse Screening for Exact Data Reduction · Sparse Screening for Exact Data Reduction ... Sparse...

Date post: 24-May-2018
Category:
Upload: duongtu
View: 222 times
Download: 0 times
Share this document with a friend
40
Center for Evolutionary Medicine and Informatics Sparse Screening for Exact Data Reduction Jieping Ye Arizona State University 1 Joint work with Jie Wang and Jun Liu
Transcript

Center for Evolutionary Medicine and Informatics

Sparse Screening for Exact Data Reduction

Jieping Ye Arizona State University

1

Joint work with Jie Wang and Jun Liu

Center for Evolutionary Medicine and Informatics

2

wide data

tall data

Center for Evolutionary Medicine and Informatics

How to do exact data reduction? The model learnt from the reduced data is identical to the model learnt from the full data:

q Lasso for wide data (feature reduction) q SVM for tall data (sample reduction)

3

Center for Evolutionary Medicine and Informatics

           

           

           

           

4

Center for Evolutionary Medicine and Informatics

Lasso/Basis Pursuit (Tibshirani, 1996, Chen, Donoho, and Saunders, 1999)

… = × +

y A z

n×1 n×p n×1

p×1

x

5

Simultaneous feature selection and regression

Center for Evolutionary Medicine and Informatics

Imaging Genetics (Thompson et al. 2013)

6

Center for Evolutionary Medicine and Informatics

Sparse Reduced-Rank Regression

7 Vounou et al. (2010, 2012)

Center for Evolutionary Medicine and Informatics

Structured Sparse Models

8

           

           

           

           

Group Lasso

Tree Lasso

Fused Lasso

Graph Lasso

Center for Evolutionary Medicine and Informatics

9

           

Sparsity has become an important modeling tool in genomics, genetics, signal and audio processing, image processing, neuroscience (theory of sparse coding), machine learning, statistics …

Center for Evolutionary Medicine and Informatics

Optimization Algorithms

•  Coordinate descent •  Subgradient descent •  Augmented Lagrangian Method •  Gradient descent •  Accelerated gradient descent •  …

10

min loss(x) + λ×penalty(x)

Center for Evolutionary Medicine and Informatics

Lasso

Fused Lasso

Group Lasso

Sparse Group Lasso

Tree Structured Group Lasso

Overlapping Group Lasso

Sparse Inverse Covariance Estimation

Trace Norm Minimization

http://www.public.asu.edu/~jye02/Software/SLEP/ 11

Center for Evolutionary Medicine and Informatics

More Efficiency?

12

Very high dimensional data

Non-smooth sparsity-induced norms

Multiple runs in model selection

A large number of runs in permutation test

Center for Evolutionary Medicine and Informatics

How to make any existing Lasso solver much more efficient?

13

Center for Evolutionary Medicine and Informatics

14

1M 1K

Data Reduction/Compression

original data reduced data

Center for Evolutionary Medicine and Informatics

Data Reduction •  Heuristic-based data reduction

–  Sure screening, random projection/selection –  Resulting model is an approximation of the true

model

•  Propose data reduction methods –  Exact data reduction via sparse screening

•  The model based on reduced data is identical to the

one constructed from complete data

15

Center for Evolutionary Medicine and Informatics

16

with screening

same solution

1M

1M 1K

without screening

Sparse Screening

Center for Evolutionary Medicine and Informatics

Large-Scale Sparse Screening

Center for Evolutionary Medicine and Informatics

Screening Rule: Motivation

Center for Evolutionary Medicine and Informatics

Large-Scale Sparse Screening (Cont’d)

Center for Evolutionary Medicine and Informatics

More on the Dual Formulation

•  Solving the dual formulation is difficult

•  Providing a good (not exact) estimate of the optimal dual solution is easier

•  A good estimate of the optimal dual solution is sufficient for effective feature screening

20

Center for Evolutionary Medicine and Informatics

Screening Rule

21

Center for Evolutionary Medicine and Informatics

Sketch of Sparse Screening

22

Center for Evolutionary Medicine and Informatics

How to Estimate the Region Θ?

J. Wang et al. NIPS’13; J. Liu et al. ICML’14

Non-expansiveness:

Center for Evolutionary Medicine and Informatics

24

Results on MNIST along a sequence of 100 parameter values along the λ/λmax scale from 0.05 to 1. The data matrix is of size 784x50,000

Center for Evolutionary Medicine and Informatics

25

Evaluation on MNIST solver   SAFE   DPP   EDPP   SDPP  

time  (s)   2245.26  685.12   233.85   45.56   9.34  

0 50 100 150 200 250 300

SAFE  DPP  EDPP  SDPP  

Speedup  

Center for Evolutionary Medicine and Informatics

Evaluation on ADNI

•  Problem: GWAS to MRI ROI prediction (ADNI) –  The size of the data matrix is 747 by 504095

Method ROI3 ROI8 ROI30 ROI69 ROI76 ROI83 Lasso Solver 37975.31 37097.25 38258.72 36926.81 38116.29 37251.03 SR 84.06 84.44 84.70 83.09 82.76 85.39 SR+Lasso 217.08 215.90 223.39 214.36 212.04 211.57 EDDP 43.56 45.75 45.70 45.01 44.31 44.16 EDDP+Lasso 183.64 190.43 182.87 170.71 177.41 178.98

Running time (in seconds) of the Lasso solver, strong rule (Tibshriani et al, 2012), and EDPP. The parameter sequence contains 100 values along the log λ/λmax scale from 100 log 0.95 to log 0.95.

Center for Evolutionary Medicine and Informatics

Sparse Screening Extensions •  Group Lasso

–  J Wang, J Liu, J Ye. Efficient Mixed-Norm Regularization: Algorithms and Safe Screening Methods. arXiv preprint arXiv:1307.4156.

•  Sparse Logistic Regression –  J Wang, J Zhou, P Wonka, J Ye. A Safe Screening Rule for Sparse Logistic

Regression. arXiv preprint arXiv:1307.4145.

•  Sparse Inverse Covariance Estimation –  S Huang, J Li, L Sun, J Liu, T Wu, K Chen, A Fleisher, E Reiman, J Ye. Learning

brain connectivity of Alzheimer’s disease by exploratory graphical models. NeuroImage 50, 935-949.

–  Witten, Friedman and Simon (2011), Mazumder and Hastie (2012)

•  Multiple Graphical Lasso –  S Yang, Z Pan, X Shen, P Wonka, J Ye. Fused Multiple Graphical Lasso. arXiv

preprint arXiv:1209.2139. 27

Center for Evolutionary Medicine and Informatics

Wide versus Tall Data

28

wide data

tall data

Center for Evolutionary Medicine and Informatics

Support Vector Machines •  SVM  is  a  maximum  margin  classiCier.

29

denotes  +1  

denotes  -­‐1  

Margin  

Center for Evolutionary Medicine and Informatics

Support Vectors •  SVM  is  determined  by  the  so-­‐called  support  vectors.

30

Support  Vectors  are  those  data  points  that  the  margin  pushes  up  against  

denotes  +1  

denotes  -­‐1  

The  non-­‐support  vectors  are  irrelevant  to  the  classiCier.  

Can  we  make  use  of  this  observation?  

Center for Evolutionary Medicine and Informatics

The Idea of Sample Screening

31

Original  Problem Screening Smaller  Problem  to  Solve

Center for Evolutionary Medicine and Informatics

Guidelines for Sample Screening

32 J. Wang, P. Wonka, and J. Ye. ICML’14.

Center for Evolutionary Medicine and Informatics

Relaxed Guidelines

33

Center for Evolutionary Medicine and Informatics

Sketch of SVM Screening

34

Center for Evolutionary Medicine and Informatics

Synthetic Studies

35

•  We  use  the  rejection  rates  to  measure  the  performance  of  the  screening  rules,  the  ratio  of  the  number  of  data  instances  whose  membership  can  be  identiCied  by  the  rule  to  the  total  number  of  data  instances.

Center for Evolutionary Medicine and Informatics

Performance of DVI for SVM on Real Data Sets

36

Comparison  of  SSNSV  (Ogawa  et  al.,  ICML’13),  ESSNSV  and  DVIs  for  SVM  on  three  real  data  sets.

IJCNN, , Speedup

Solver Total 4669.14

Solver + SSNSV

SSNSV 2.08

2.31 Init. 92.45

Total 2018.55

Solver + ESSNSV

ESSNSV 2.09

3.01 Init. 91.33

Total 1552.72

Solver + DVI

DVI 0.99

5.64 Init. 42.67

Total 828.02

Wine, , Speedup

Solver Total 76.52

Solver + SSNSV

SSNSV 0.02

3.50 Init. 1.56

Total 21.85

Solver + ESSNSV

ESSNSV 0.03

4.47 Init. 1.60

Total 17.17

Solver + DVI

DVI 0.01

6.59 Init. 0.67

Total 11.62

Covertype, , Speedup

Solver Total 1675.46

Solver + SSNSV

SSNSV 2.73

7.60 Init. 35.52

Total 220.58

Solver + ESSNSV

ESSNSV 2.89

10.72 Init. 36.13

Total 156.23

Solver + DVI

DVI 1.27

79.18 Init. 12.57

Total 21.26

Center for Evolutionary Medicine and Informatics

Experiments on Real Data Sets

37

Comparison  of  SSNSV  (Ogawa  et  al.,  ICML’13),  ESSNSV  and  DVIs  for  LAD  on  three  real  data  sets.

Telescope, , Speedup

Solver Total 122.34

Solver + DVI

DVI 0.28

9.86 Init. 0.12

Total 12.14

Computer, , Speedup

Solver Total 5.85

Solver + DVI

DVI 0.08

19.21 Init. 0.05

Total 0.28

Telescope, , Speedup

Solver Total 21.43

Solver + DVI

DVI 0.06

114.91 Init. 0.1

Total 0.19

Center for Evolutionary Medicine and Informatics

Summary •  Developed exact data reduction approaches

–  Exact data reduction via feature screening –  Exact data reduction via sample screening

•  The model based on reduced data is identical to the one constructed from complete data

•  Results show screening leads to a significant speedup.

•  Extend exact data reduction to other sparse learning formulations –  Sparsity on features, samples, networks etc

38

Center for Evolutionary Medicine and Informatics

Resource

39

•  Tutorial  webpages  of  our  screening  rules,  which  include  sample  codes,  implementation  instructions,  illustration  materials,  etc.  

http://www.public.asu.edu/~jwang237/screening.html

Seven  lines  implementation  of  EDPP  rule  

The  list  is  growing  quickly  

Center for Evolutionary Medicine and Informatics

40


Recommended