Lab 1

Getting started with CLOPand

the Spider package

CLOP Tutorial

• CLOP=Challenge Learning Object Package.

• Based on the Spider developed at the Max Planck Institute.

• Two basic abstractions:– Data object– Model object

http://clopinet.com/isabelle/Projects/modelselect/MFAQ.html

CLOP Data Objects

cd <code_dir>use_spider_clop;X=rand(10,8);Y=[1 1 1 1 1 -1 -1 -1 -1 -1]';D=data(X,Y); % constructor[p,n]=get_dim(D) get_x(D) get_y(D)

At the Matlab prompt:

CLOP Model Objects

model = kridge; % constructor[resu, model] = train(model, D);resu, model.W, model.b0Yhat = D.X*model.W' + model.b0testD = data(rand(3,8), [-1 -1 1]');tresu = test(model, testD); balanced_errate(tresu.X, tresu.Y)

D is a data object previously defined.

Hyperparameters and Chains

default(kridge) hyper = {'degree=3', 'shrinkage=0.1'}; model = kridge(hyper);

model = chain({standardize,kridge(hyper)}); [resu, model] = train(model, D); tresu = test(model, testD); balanced_errate(tresu.X, tresu.Y)

A model often has hyperparameters:

Models can be chained:

Hyper-parameters

• http://clopinet.com/isabelle/Projects/modelselect/MFAQ.html

• Kernel methods: kridge and svc:k(x, y) = (coef0 + x y)degree exp(-gamma ||x - y||2)

kij = k(xi, xj)

kii kii + shrinkage

• Naïve Bayes: naive: none• Neural network: neural

units, shrinkage, maxiter

• Random Forest: rf (windows only)mtry

Lab 2

Getting started with the NIPS 2003 FS challenge

The Datasets

• Arcene: cancer vs. normal with mass-spectrometry analysis of blood serum.

• Dexter: filter texts about corporate acquisition from Reuters collection.

• Dorothea: predict which compounds bind to Thrombin from KDD cup 2001.

• Gisette: OCR digit “4” vs. digit “9” from NIST.

• Madelon: artificial data.

http://clopinet.com/isabelle/Projects/NIPS2003/Slides/NIPS2003-Datasets.pdf

Data Preparation

• Preprocessing and scaling to numerical range 0 to 999 for continuous data and 0/1 for binary data.

• Probes: Addition of “random” features distributed similarly to the real features.

• Shuffling: Randomization of the order of the patterns and the features.

• Baseline error rates (errate): Training and testing on various data splits with simple methods.

• Test set size: Number of test examples needed using rule-of-thumb ntest = 100/errate.

Data Statistics

Dataset Size Type FeaturesTraining Examples

Validation Examples

Test Examples

Arcene8.7 MB

Dense 10000 100 100 700

Gisette22.5 MB

Dense 5000 6000 1000 6500

Dexter0.9 MB

Sparse integer

20000 300 300 2000

Dorothea4.7 MB

Sparse binary

100000 800 350 800

Madelon2.9 MB

Dense 500 2000 600 1800

ARCENE

• Sources: National Cancer Institute (NCI) and Eastern Virginia Medical School (EVMS).

• Three datasets: 1 ovarian cancer, 2 prostate cancer, all preprocessed similarly.

• Task: Separate cancer vs. normal.

ARCENE is the cancer dataset

0 2000 4000 6000 8000 10000 12000 14000 160000

10

20

30

40

50

60

70

80

90

100

DEXTER

• Sources: Carnegie Group, Inc. and Reuters, Ltd.

• Preprocessing: Thorsten Joachims.• Task: Filter “corporate acquisition” texts.

DEXTER filters texts

NEW YORK, October 2, 2001 – Instinet Group Incorporated (Nasdaq: INET), the world’s largest electronic agency securities broker, today announced that it has completed the acquisition of ProTrader Group, LP, a provider of advanced trading technologies and electronic brokerage services primarily for retail active traders and hedge funds. The acquisition excludes ProTrader’s proprietary trading business. ProTrader’s 2000 annual revenues exceeded $83 million.

DOROTHEA

• Sources: DuPont Pharmaceuticals Research Laboratories and KDD Cup 2001.

• Task: Predict compounds that bind to Thrombin.

DOROTHEA is the Thrombin dataset

GISETTE

• Source: National Institute of Standards and Technologies (NIST).

• Preprocessing: Yann LeCun and collaborators.

• Task: Separate digits “4” and “9”.

GISETTE contains handwritten digits

5 10 15 20 25

5

10

15

20

25

5 10 15 20 25

5

10

15

20

25

MADELON

• Source: Isabelle Guyon, inspired by Simon Perkins et al.

• Type of data: Clusters on the summits of a hypercube.

MADELON is random data

Performance Measures

Prediction

Class -1 Class +1

Truth Class -1 a b

Class +1 c d

Confusion matrix

• Balanced Error Rate (BER): the average of the error rates for each class: BER = 0.5*(b/(a+b) + c/(c+d)).

• Area Under Curve (AUC): the area under the ROC curve obtained by plotting a/(a+b) against d/(c+d) for each confidence value, starting at (0,1) and ending at (1,0).

• Fraction of Features (FF): the ratio of the num. of features selected to the total num. of features in the dataset.

• Fraction of Probes (FP): the ratio of the num. of “garbage features” (probes) selected to the total num. of feat. select.

BER distribution

0 5 10 15 20 25 30 35 40 45 500

20

40ARCENE

0 5 10 15 20 25 30 35 40 45 500

20

40

DEXTER

0 5 10 15 20 25 30 35 40 45 500

20

40

DOROTHEA

0 5 10 15 20 25 30 35 40 45 500

20

40GISETTE

0 5 10 15 20 25 30 35 40 45 500

20

40

MADELON

Test error (%)

Power of Feature Selection

Best frac. feat

Actual frac. probes

ARCENE 5% 30%

DEXTER 1.5% 50%

DOROTHEA 0.3% 50%

GISETTE 18% 50%

MADELON 1.6% 96%

Visualization

1) Create a heatmap of the data matrix:show(D.train);

2) Look at individual patterns: browse(D.train);

3) Make a scatter plot of the 2 first features:show(D.train);

4) Visualize the results:[Dat,Model]=train(model, D.train);Dat=test(Model, D.valid);roc(Dat);

BER = f(threshold)

Theta = -37.40

Training set

Theta = -38.14

Test set

No bias adjustment, test BER=22.54%; with bias, test BER=12.37%

-100 0 100 200 300 400 5000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

threshold

-100 -50 0 50 100 150 200 250 300 350 4000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

threshold

sensitivity sensitivity

specificity specificity

BER BER

DOROTHEA

ROC curve

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 10

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

AUC=0.91

Specificity

Sen

sitiv

ityDOROTHEA

Feature SelectionMADELON (pval_max=0.1)

0 10 20 30 40 50 60 70 80 901

1.05

1.1

1.15

1.2

1.25

1.3

1.35

1.4

rank

W

0 10 20 30 40 50 60 70 80 900

0.1

0.2

0.3

0.4

0.5

0.6

0.7

rank

FD

R

Heat mapARCENE

5 10 15 20 25 30 35 40 45 50

10

20

30

40

50

60

70

80

90

100 0

100

200

300

400

500

600

700

800

Scatter plots

chain({standardize, s2n('f_max=2'), normalize, my_svc})

Test BER=49%

0 50 100 150 2000 20 40 60 80 100

0

50

100

150

200

0

20

40

60

80

100

0 100 200 300 4000 100 200 300 400

0

100

200

300

400

0

100

200

300

400

chain({standardize, s2n('f_max=1100'), normalize, gs('f_max=2'), my_svc})

Test BER=29.37%

ARCENE

Lab 3

Playing with FS filters and classifiers on Madelon and Dexter

Lab 3 software

1) Try the examples in Lab3 README.m2) Inspiring your self by the examples,

write a new feature ranking filter object. Choose one in Chapter 3 or invent your own.

3) Provide the pvalue and FDR (using a tabulated distribution or the probe method).

Filters: see chapter 3

Filters Implemented

• @s2n• @Relief• @Ttest• @Pearson (Use Matlab corrcoef. Gives the same results

as Ttest, classes are balanced.)

• @Ftest (gives the same results as Ttest. Important for the pvalues: the Fisher criterion needs to be multiplied by num_patt_per_class or use anovan.)

• @aucfs (ranksum test)

Evalution of pval and FDR

• Ttest object: – computes pval analytically– FDR~pval*nsc/n

• probe object: – takes any feature ranking object as an

argument (e.g. s2n, relief, Ttest)– pval~nsp/np

– FDR~pval*nsc/n

Analytic vs. probe

0 500 1000150020002500300035004000450050000

0.02

0.04

0.06

0.08

0.1

0.12

0.14

rank

FD

R

Arcene

0 500 1000150020002500300035004000450050000

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

rank

FD

R

Dexter

0 5001000150020002500300035004000450050000

1

2

x 10-4

rank

FD

R

Dorothea

0 500 1000150020002500300035004000450050000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

rank

FD

R

Gisette

0 50 100 150 200 250 300 350 400 450 5000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

rank

FD

RMadelon

Red analytic – Blue probe

Relief vs. Ttest (Madelon)

0 5 10 15 20 25 30 35 40 45 500

0.01

0.02

0.03

0.04

0.05

0.06

0.07

0.08

0.09

rank

pva

l

0 5 10 15 20 25 30 35 40 45 500

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

FD

R

Ttest

Relief Ttest

Relief

rank

Lab 4

Plying with Feature construction

on Gisette

Gisette

• Handwritten digits.• Goal: get familiar with the data and result

formats. Make a first submission.• Easiest LM for Gisette: naive and svc.• Best preprocessing: normalize.• Easiest feature selection method: s2n.• Many training examples (6000). Unsuitable

for kridge unless subsampling is used.• Many features (5000). Select features before

running neural or rf.

Baseline Model

baselineGisette (BER=1.8%, feat=20%)

my_classif=svc({'coef0=1', 'degree=3', 'gamma=0', 'shrinkage=1'});

my_model=chain({normalize, s2n('f_max=1000'), my_classif});

Baseline methods

baselineGisette (CV=1.91%, test=1.80%, feat=20%)my_classif=svc({'coef0=1', 'degree=3', 'gamma=0', 'shrinkage=1'});

my_model=chain({normalize, s2n('f_max=1000'), my_classif});

baselineGisette2 (CV=1.34%, test=1.17%, feat=20%)my_model=chain({s2n('f_max=1000'), normalize, my_classif});

pixelGisette (CV=1.31%, test=0.91%)my_classif=svc({'coef0=1', 'degree=4', 'gamma=0', 'shrinkage=0.1'});

my_model=chain({normalize, my_classif});

ConvolutionsGISETTE (pixelGisette_exp_conv)Exponential kernel, sigma1=1.8 , sigma2=1.8

1 2 3 4 5 6 7 8 9

1

2

3

4

5

6

7

8

9

prepro=my_model{1};

show(prepro.child);

DD=test(prepro,D.train);

browse_digit(DD.X, D.train.Y); Index: 4 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 4 Class: 1

5 10 15 20 25

5

10

15

20

25

chain({convolve(exp_ker({'dim1=9', 'dim2=9'})), normalize, my_classif})

Principal Components

Index: 1 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 2 Class: -1

5 10 15 20 25

5

10

15

20

25

Index: 3 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 4 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 5 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 6 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 7 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 8 Class: 1

5 10 15 20 25

5

10

15

20

25

@pca_bank: Filter bank object retaining the first f_max principal components of the data matrix

@kmeans_bank: Filter bank containing templates corresponding to f_max cluster centers.

Hadamard bank

10 20 30 40 50 60 70 80

10

20

30

40

50

60

70

@hadamard_bank: Filter bank object performing a Hadamard transform.

Hadamard bank

Fourier Transform

Index: 5 Class: 1

5 10 15 20 25

5

10

15

20

25

Index: 5 Class: 1

5 10 15 20 25

5

10

15

20

25

@fourier: Two dimensional Fourier transform.

Epilogue

Becoming a pro andplaying with

other datasets

Baseline Methods for the Feature Extraction Class Isabelle Guyon

GISETTE Best BER=1.26Best BER=1.260.14% - n0=1000 (20%) – BER0=1.80%0.14% - n0=1000 (20%) – BER0=1.80%

my_classif=svc({'coef0=1', 'degree=3', 'gamma=0', 'shrinkage=1'});

my_model=chain({normalize, s2n('f_max=1000'), my_classif});

ARCENE Best BER= 11.9 Best BER= 11.9 1.2 %1.2 % - n0=1100 (11%) – BER0=14.7%- n0=1100 (11%) – BER0=14.7%

my_svc=svc({'coef0=1', 'degree=3', 'gamma=0', 'shrinkage=0.1'});

my_model=chain({standardize, s2n('f_max=1100'), normalize, my_svc})

BACKGROUND

We present supplementary course material complementing the book “Feature Extraction, Fundamentals and Applications”, I. Guyon et al Eds., to appear in Springer. Classical algorithms of feature extraction were reviewed in class. More attention was

given to the feature selection than feature construction because of the recent success of methods involving a large number of "low-level" features.

The book includes the results of a NIPS 2003 feature selection challenge. The students learned techniques employed by the best challengers and tried to match the best performances. A Matlab® toolbox was provided with sample code.The students could makepost-challenge entries to: http://www.nipsfsc.ecs.soton.ac.uk/.

Challenge: Good performance + few features. Tasks: Two-class classification. Data split: Training/validation/test. Valid entry: Results on all 5 datasets.

DATASETS

Dataset Size MB Type Features Training Validation Test

Arcene 8.7 Dense 10000 100 100 700

Gisette 22.5 Dense 5000 6000 1000 6500

Dexter 0.9 Sparse 20000 300 300 2000

Dorothea 4.7 Sparse 100000 800 350 800

Madelon 2.9 Dense 500 2000 600 1800

MADELON Best BER=6.22Best BER=6.220.57% - n0=20 (4%) – BER0=7.33%0.57% - n0=20 (4%) – BER0=7.33%

my_classif=svc({'coef0=1', 'degree=0', 'gamma=1', 'shrinkage=1'});

my_model=chain({probe(relief,{'p_num=2000', 'pval_max=0'}), standardize, my_classif})

METHODS

Scoring:• Ranking according to test set balanced error rate (BER) , i.e. the average positive class error rate and negative class error rate. • Ties broken by the feature set size.

Learning objects:• CLOP learning objects implemented in Matlab®.• Two simple abstractions: data and algorithm.• Download: http://www.modelselect.inf.ethz.ch/models.php.

Task of the students:• Baseline method provided, BER0 performance and n0 features.• Get BER<BER0 or BER=BER0 but n<n0.• Extra credit for beating the best challenge entry.• OK to use the validation set labels for training.

DOROTHEA Best BER=8.54Best BER=8.540.99% - n0=1000 (1%) – BER0=12.37%0.99% - n0=1000 (1%) – BER0=12.37%

my_model=chain({TP('f_max=1000'), naive, bias});

DEXTER Best BER=3.30Best BER=3.300.40% - n0=300 (1.5%) – BER0=5%0.40% - n0=300 (1.5%) – BER0=5%

my_classif=svc({'coef0=1', 'degree=1', 'gamma=0', 'shrinkage=0.5'});

my_model=chain({s2n('f_max=300'), normalize, my_classif})

ARCENE

0 5 10 15 20 25 30 35 40 45 50

DEXTER

0 5 10

15

20

25

30

35

40

45

50

DEXTER: text categorization

GISETTE

0 5 10 15 20 25 30 35 40 45 50

GISETTE: digit recognition

DOROTHEA: drug discoveryDOROTHEA

0 5 10

15

20

25

30

35

40

45

50

RESULTS

0 5 10 15 20 25 30 35 40 45 50

MADELON

MADELON: artificial data

CONCLUSIONSThe performances of the challengers could be matched with the CLOP library. Simple filter methods (S2N and Relief) were sufficient to get a space dimensionality reduction comparable to what the winners obtained. SVMs are easy to use and generally work better than other methods. We experienced with Gisette to add prior knowledge about the task and could outperform the winners. Further work includes using prior knowledge for other datasets.

5 10 15 20 25

5

10

15

20

25

5 10 15 20 25

5

10

15

20

25

NEW YORK, October 2, 2001 – Instinet Group Incorporated (Nasdaq: INET), the world’s largest electronic agency securities broker, today announced tha

ARCENE: cancer diagnosis0 2000 4000 6000 8000 10000 12000 14000 16000

0

10

20

30

40

50

60

70

80

90

100

Best student results

http://clopinet.com/isabelle/Projects/ETH/Feature_Selection_w_CLOP.html

Agnostic Learning vs.

Prior Knowledge challenge

Isabelle Guyon, Amir Saffari, Gideon Dror, Gavin Cawley, Olivier Guyon,

and many other volunteers, see http://www.agnostic.inf.ethz.ch/credits.php

Open until August 1st, 2007

Datasets

Dataset Domain Type Feat-ures

Training Examples

Validation Examples

Test Examples

ADA Marketing Dense 48 4147 415 41471

GINA Digits Dense 970 3153 315 31532

HIVADrug discovery

Dense 1617 3845 384 38449

NOVAText classif.

Sparse binary 16969 1754 175 17537

SYLVA Ecology Dense 216 13086 1308 130858

http://www.agnostic.inf.ethz.ch

ADAADA is the marketing database

• Task: Discover high revenue people from census data. Two-class pb.

• Source: Census bureau, “Adult” database from the UCI machine-learning repository.

• Features: 14 original attributes including age, workclass,  education, education, marital status, occupation, native country. Continuous, binary and categorical features.

•  

GINA

• Task: Handwritten digit recognition. Separate the odd from the even digits. Two-class pb. with heterogeneous classes.

• Source: MNIST database formatted by LeCun and Cortes.

• Features: 28x28 pixel map.

•  

GINA is the digit database

HIVAHIVA is the HIV database

• Task: Find compounds active against the AIDS HIV infection. We brought it back to a two-class pb. (active vs. inactive), but provide the original labels (active, moderately active, and inactive).

• Data source: National Cancer Inst.• Data representation: The compounds are

represented by their 3d molecular structure.•  

NOVANOVA is the text classification database

• Task: Classify newsgroup emails into politics or religion vs. other topics.

• Source: The 20-Newsgroup dataset from in the UCI machine-learning repository.

• Data representation : The raw text with an estimated 17000 words of vocabulary.

Subject: Re: Goalie masksLines: 21

Tom Barrasso wore a great mask, one time, last season.  He unveiled it at a game in Boston. 

It was all black, with Pgh city scenes on it. The "Golden Triangle" graced the top, alongwith a steel mill on one side and the Civic Arena on the other.   On the back of the helmet was the old Pens' logo the current (at the time) Penslogo, and a space for the "new" logo.

A great mask done in by a goalie's superstition.

Lori 

SYLVASYLVA is the ecology database

• Task: Classify forest cover types into Ponderosa pine vs. everything else.

• Source: US Forest Service (USFS). • Data representation: Forest cover type for 30

x 30 meter cells encoded with 108 features (elavation, hill shade, wilderness type, soil type, etc.)

•  

BER distribution (March 1st)

Agnostic learning Prior knowledge

The black vertical line indicates the best ranked entry (only the 5 last entry of each participant were ranked). Beware of overfitting!

CLOP models

Preprocessing and FS

Model grouping

for k=1:10

base_model{k}=chain({standardize, naive});

end

my_model=ensemble(base_model);

CLOP models (best entrant)

 

Dataset CLOP models selected

ADA 2*{sns,std,norm,gentleboost(neural),bias}; 2*{std,norm,gentleboost(kridge),bias}; 1*{rf,bias}

GINA 6*{std,gs,svc(degree=1)}; 3*{std,svc(degree=2)}

HIVA 3*{norm,svc(degree=1),bias}

NOVA 5*{norm,gentleboost(kridge),bias}

SYLVA 4*{std,norm,gentleboost(neural),bias}; 4*{std,neural}; 1*{rf,bias}

 

Juha Reunanen, cross-indexing-7

sns = shift’n’scale, std = standardize, norm = normalize (some details of hyperparameters not shown)

CLOP models (2nd best entrant)

 

Dataset CLOP models selected

ADA {sns, std, norm, neural(units=5), bias}

GINA {norm, svc(degree=5, shrinkage=0.01), bias}

HIVA {std, norm, gentleboost(kridge), bias}

NOVA {norm,gentleboost(neural), bias}

SYLVA {std, norm, neural(units=1), bias}

 

Hugo Jair Escalante Balderas, BRun2311062

sns = shift’n’scale, std = standardize, norm = normalize (some details of hyperparameters not shown)

Note: entry Boosting_1_001_x900 gave better results, but was older.