This is a place for instructor and students to record difficulties encountered during the course.

Resources

Course Homepage
NIPS 2003 Feature Selection Challenge
The results of post challenge entries are here:
- overall
- arcene
- gisette
- dexter
- dorothea
- madelon

Tasks

pipibjc is going to write a matlab2libsvm.m

Progress

Week 12

One group starts to study others' approach

I have submit the overall average (3.60429, similar to 3.605 in the subset) to Netflix. The leaderboard will only show results that outperform Cinematch (<0.9514), but the RMSE of quiz subset will be sent by email.

The RMSE of overall-average prediction is 1.1309.

Week 10

zao studies KDD Cup 2007 rules

Week 9

pipibjc studies how to use rmse.pl
two groups study two proposed approach

Discussions

zao group proposes k-means to grouping users then tries nmf
- take movie quality into consideration
ma group proposes the following
- movie classification to generate missing values in the matrix, so we can calculate the distance measure
- use people's favor to refine movie classification results (active learning) (optional)
- kNN

Discussion of Feature Selection Competition

DOROTHEA: setting weighted penalty term helps

Our Results on Feature Selection Competition

Result Matrix

student	ARCENE	DEXTER	DOROTHEA	GISETTE	MADELON
ma	9.45	3.65	10.51	1.32	6.28
oyster	13.16	4.1	8.82	1.32	6.72
pipibjc	10.22	3.80	12.05	1.43	6.83
zao	10.12	3.65	12.04	1.09	6.89

Approach

ARCENE

student	approach
ma	`my_svc=svc({'coef0=6.069413', 'degree=3', 'gamma=0', shrinkage=0.1'}); my_model=chain({relief('f_max=1378'),normalize,my_svc});`
oyster	`my_svc=svc({'coef0=2', 'degree=0', 'gamma=1.0', 'shrinkage=0.1'}); my_model=chain({normalize, pc_extract('f_max=130'), my_svc});`
pipibjc	`my_svc=svc({'coef0=2','degree=1','gamma=3','shrinkage=0.1'}); relief_chain=chain({normalize,relief({'f_max=1400'})}); s2n_chain=chain({standardize,s2n({'f_max=900'})}); feat_sel=combine_feat({relief_chain,s2n_chain}); my_model=chain({feat_sel,normalize,my_svc});`
zao	`my_svc=svc({'coef0=2', 'degree=3', 'gamma=0', 'shrinkage=0.1'}); my_model=chain({relief('f_max=1600'),normalize,relief('f_max=1400'),normalize,my_svc});`

DEXTER

student	approach
ma	`my_svc=svc({'coef0=1.702212', 'degree=2', 'gamma=0', 'shrinkage=0.5'}); my_model=chain({s2n('f_max=2185'),normalize,my_svc});`
oyster	`my_svc=svc({'coef0=0', 'degree=0', 'gamma=0.9', 'shrinkage=0.1'}); my_model=chain({relief('f_max=3425'), normalize, my_svc });`
pipibjc	`my_svc=svc({'coef0=1','degree=1','gamma=0','shrinkage=0.5'}); relief_chain=relief({'f_max=2500'}); s2n_chain=s2n({'f_max=4500'}); feat_sel=combine_feat({relief_chain,s2n_chain}); my_model=chain({normalize,feat_sel,my_svc});`
zao	`my_classif=svc({'coef0=1', 'degree=1', 'gamma=0', 'shrinkage=0.5'}); my_model=chain({s2n('f_max=6000'), normalize, s2n('f_max=4500'), normalize, s2n('f_max=2500'), normalize, my_classif});`

DOROTHEA

student	approach
ma	`my_svc=svc({'coef0=0', 'degree=0', 'gamma=4.2', 'shrinkage=0.05', 'C=1'}); my_model=chain({TP('f_max=15000'), normalize, relief({'f_max=715','k_num=28'}),my_svc, bias});`
oyster	`my_classif=svc({'coef0=1', 'degree=0', 'gamma=2.6', 'C=0.25', 'w1=10'}); my_model=chain({TP('f_max=15000'), normalize, relief('f_max=700'), my_classif, bias});`
pipibjc	`s2n_chain=chain({normalize,s2n({'f_max=1500'})}); my_model=chain({s2n_chain,naive,bias});`
zao	`my_model=chain({normalize,TP('f_max=15000'),normalize,relief('f_max=700'),naive, bias});`

GISETTE

student	approach
ma	`my_svc=svc({'coef0=0', 'degree=0', 'gamma=0.87', 'shrinkage=0.7'}); my_model=chain({s2n('f_max=1000'), pc_extract('f_max=50'), normalize, my_svc});`
oyster	`my_classif=svc({'coef0=1', 'degree=3', 'gamma=0', 'shrinkage=1'}); my_model=chain({ relief('f_max=1940'), normalize, relief('f_max=1000'), normalize, my_classif});`
pipibjc	`my_model=chain({relief('f_max=1000'),normalize,knn({'k=5'})});`
zao	`my_classif=svc({'coef0=1', 'degree=3', 'gamma=0', 'shrinkage=1'}); my_model=chain({relief('f_max=2000'), s2n('f_max=1000'), normalize, my_classif});`

MADELON

student	approach
ma	`my_svc=svc({'coef0=0', 'degree=0', 'gamma=0.2091', 'shrinkage=0.5741'}); my_model=chain({relief({'f_max=20'}), standardize, my_svc});`
oyster	`my_classif=svc({'coef0=1', 'degree=0', 'gamma=0.3', 'shrinkage=0.5'}); my_model=chain({relief('f_max=210'), normalize, relief('f_max=20'), standardize, my_classif});`
pipibjc	`my_svc=svc({'coef0=1','degree=2','gamma=0.3','shrinkage=0.3'}); relief_chain=chain({normalize,relief({'f_max=20'})});`
zao	`my_classif=svc({'C=1', 'gamma=0.35', 'degree=0', 'shrinkage=1'}); my_model=chain({relief('f_max=100'), normalize, relief('f_max=20'), standardize, my_classif});`

Week 8

Problems

oyster claims using only training data is better (first two sets)
ma claims that for DOROTHEA, you need bias + TP
For DOROTHEA, only training is used

Week 7

Tasks

refine your submission
submit the rest three datasets

Problems

Why RBF kernel tends to predict majority class when gamma goes to infinity?

Week 5

Tasks

randomNN
What exactly probe does?
Each one submits one prediction for arcene+dexter

Problems

RandomForest is windows only
gridsel issues

Week 4

Tasks

Why can we use full tree in random forest?
Discuss your strategy?

Problems

Week 3

Tasks

study five preprocessors (zao), five additional feature selection methods (oyster), five additional feature selection methods (Ma, pipibjc)
how they come up with best entry, for example higher validation ber ⇒ higher test ver

Problems

missing range.m (from Statistics toolbox)

function y = range(x)
  y = max(x)-min(x);

We need DataNCode.zip from the link on their page
Some BER values on Table 3 is different from Score/Simple-22-Apr-2006.score (DataNCode.zip). For example, DEXTER or GISETTE.

Week 2

Tasks

able to reproduce their baseline and best entries

Problems

LIBSVM on Linux REASON

--- Makefile_orig.orig  2005-05-12 20:28:32.000000000 +0800
+++ Makefile_orig       2007-03-03 20:07:22.000000000 +0800
@@ -1,11 +1,12 @@
 # This Makefile is used under Linux

 MATLABDIR ?= /usr/local/matlab
-CXX = g++
+CXX = g++-3.3
 CFLAGS = -Wall -O3 -fPIC -I$(MATLABDIR)/extern/include

 MEX = $(MATLABDIR)/bin/mex
 MEX_OPTION = CC\#$(CXX) CFLAGS\#"-Wall -O3 -fPIC"
+MEX_OPTION += -largeArrayDims

 all:    svmpredict svmtrain
--- svm_model_matlab.c.orig     2005-04-25 19:54:28.000000000 +0800
+++ svm_model_matlab.c  2007-03-03 20:06:44.000000000 +0800
@@ -4,6 +4,10 @@

 #include "mex.h"

+#if MX_API_VER < 0x07030000
+typedef int mwIndex;
+#endif
+
 #define NUM_OF_RETURN_FIELD 10

 static const char *field_names[] = {
@@ -112,7 +116,8 @@

        // SVs
        {
-               int *ir, *jc, ir_index, nonzero_element;
+               int ir_index, nonzero_element;
+               mwIndex *ir, *jc;
                mxArray *pprhs[1], *pplhs[1];

                nonzero_element = 0;
@@ -256,7 +261,7 @@
        {
                int sr, sc, elements;
                int num_samples;
-               int *ir, *jc;
+               mwIndex *ir, *jc;
                mxArray *pprhs[1], *pplhs[1];

                // transpose SV

> diff -u svmtrain.c.orig svmtrain.c

--- svmtrain.c.orig     2005-09-23 13:56:22.000000000 +0800
+++ svmtrain.c  2007-03-03 20:07:47.000000000 +0800
@@ -7,6 +7,10 @@
 #include "mex.h"
 #include "svm_model_matlab.h"

+#if MX_API_VER < 0x07030000
+typedef int mwIndex;
+#endif
+
 #define CMD_LEN 2048
 #define Malloc(type,n) (type *)malloc((n)*sizeof(type))

@@ -258,7 +262,7 @@
 void read_problem_sparse(const mxArray *label_vec, const mxArray *instance_mat)
 {
     int i, j, k, low, high;
-    int *ir, *jc;
+    mwIndex *ir, *jc;
     int elements, max_index, num_samples;
     double *samples, *labels;
     mxArray *instance_mat_tr; // transposed instance sparse matrix

sample_code/main.m: remove all calls to upper()

Sitemap

Resources

Tasks

Progress

Week 13

Week 12

Week 11

Week 10

Week 9

Discussions

Discussion of Feature Selection Competition

Our Results on Feature Selection Competition

Result Matrix

Approach

Week 8

Problems

Week 7

Tasks

Problems

Week 5

Tasks

Problems

Week 4

Tasks

Problems

Week 3

Tasks

Problems

Week 2

Tasks

Problems