readme

统计学习软件包

License 

This software is free only for non-commercial use. It must not be modified
and  distributed without prior permission of the author. The author is not
responsible for implications from the use of this software. 
This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY.  
                                                     


General Description: 

This is an implementation of the Generalized LASSO method
for classification. Details can be found in 
Roth, V. (2002),  The Generalized LASSO: a wrapper approach to gene selection
for microarray data. University of Bonn, Dep. Computer Science III,
Tech.Rep. IAI-TR-2002-8. Please cite this reference if you use this program.

The user must provide a file which contains the feature values for each
sample (One feature per line, see below). Standardization for zero mean and
unit variance usually is a very good idea. 
The optimal l_1 constraint value kappa is automatically found by averaging over
randomly chosen  80%/20% training/test splits of the dataset. The prediction
performance of the finally selected model is estimated on the basis of new 
randomly chosen splits.  

A demo application for Microarray data (sample classification) is
included.

In order to obtain nonlinear variants of the Generalized LASSO, the features
may also be preprocessed by a 
Mercer kernel, resulting in a (N x N) input file. RBF-kernels usually work
best.   



Installation:

Download GenLASSO.tar.gz.
Type "tar xfvz GenLASSO.tar.gz". This will install the sources into
a directory called GenLASSO.  

The Generalized LASSO classifier uses part of the donlp2 optimization package
by P. Spellucci (to be exact the "donlp2_ansi_c.tar.gz" version rewritten in
Ansi C  by Serge Schoeffert). The latter can be downloaded from  
http://plato.la.asu.edu/donlp2.html. 
Please unzip/untar "donlp2_ansi_c.tar.gz"  and copy  "donlp2.c" and
"user_eval.c" and all header-files "*.h" into your GenLASSO directory before
running the make program. 

Edit the Makefile and eventually adjust your compiler settings.

Type "make", close your eyes, stop breathing and hope for the best. If you are
lucky, this should create the executable "GenLASSO_C". Otherwise follow your
intuition. Re-start breathing in any case. To avoid panic, however, eyes might
be kept close. 


Usage:

"GenLASSO_C file_name", where file_name specifies a configuration file.

Syntax of configuration file:

1. line: <number of samples>  <number of features>
 -> The problem dimensions

2. line: <initial constraint value> <constraint increment> <number of
increments> 
 -> specifies the grid-search for the optimal l_1 constraint value kappa. 

3. line: <Number of cross-validation iterations>
  -> Number of training/test splits of dataset used for model selection and
  assessment 

4. line: <data filename>
 -> Name of file containing both the class labels (first line, either 0 or 1)
 and the 
 (pre-processed) 
 expression levels (one line per gene). See also the example datafile
 "golub.dat" 

5. line: <annotation filename>
  -> Name of file containing feature (gene) annotations, one per line, See
  also example  file "golub_names.dat"

6. line: <reduce flag>
  -> if set to one, the model is re-trained on a subset of genes, consisting
  of those genes which have stability scores > 0.05 in the first run.

See also the included example parameter file "golub_config.dat".

Demo:

Preprocessed data from (Golub et al.) 
(www-genome.wi.mit.edu/mpr/data_set_ALL_AML.html):
Original dataset consists of 72 samples, of which 47 are acute lymphoblastic
leukemia (ALL),  and 25 samples are acute myeloid leukemia (AML). Expression
levels of 7129 genes were measured using Affymetrix high-density
oligonucleotide arrays. We preprocessed the data by firstly excluding genes
with mostly  negative intensity values, and secondly by excluding genes with
max/min <2   and max-min < 1000. This leaves us with a reduced set of 1479
genes. Finally, the data were log-transformed, "squashed" through a
tanh-function (for outlier--reduction), and standardized to zero mean and
unit variance across samples. Class labels and expression values are
summarized in file "golub.dat".

To run the demo application, type  "GenLASSO_C golub_config.dat".   
Results (including stability scores and annotations) will be summarized in
the file "relevant_genes.dat": model parameters, cross-validation error
(standard deviation), average number of "relevance vectors" (RVs), stability
scores and gene annotations.

Graphical output is provided by the file "rg.pgm". 
The error rates for thresholding by different confidence levels  and the
fraction of samples contained in the doubt class are printed in file
"doubt.dat". Format per line: <threshold> <error rate> <fraction of samples in
doubt class>.