Source code and data sets from http://www.ailab.si/supp/bmcg-replgroups/

CONTENTS

- experimental data are located in splittingvswhole/,
- scripts for the generation of synthetic data sets are in snpgen/,
- python interface to HFCC resides in hfcc/,
- faster interaction gain calculation is provided by mOrngInteract.py and intgain_install.
- the actual experimental run in is RUN.py.


INSTALLATION

A 32-bit i386 Linux system is needed to perform the analyses.

Before running them, you will need to:
1. Install Orange beta (the scripting interface is sufficient).
2. Put HFCC binary to the correct place.
3. Install a Python module for faster computation of interaction gain.

Step by step instructions for performing each of the three steps follow. 
The whole procedure was tested on Ubuntu 9.04. 

1. Installation of Orange beta (with the scripting interface enabled only)

Described at http://www.ailab.si/svn/orange/trunk/orange/INSTALL.linux.txt


2. HFCC binary

Download HFCC binaries from http://www.neocodex.com/en/hfcc.0.5.zip and put
hfcc5 file from the archive into the hfcc/ subfolder of this supplement.


3. Python module for faster computation of interaction gains

Go to intgain_install/ subfolder of this supplement and run
sudo python setup.py install


USAGE AND DESCRIPTION

To run the experiments, run the RUN.py python script. As the calculation of 
all results takes a very long time (1-2 weeks), we provided options to 
calculate only a part of the results:
- "python RUN.py MODEL1" produces results of first 10 repetitions for the 
  smaller synthetic models (takes about 5 minutes)
- "python RUN.py FIRST" produces results of the first repetitions for all 
  reported data sets (about 4 hours).

Expected times are reported for a computer with Intel Core 2 2.4GHz processor 
and 2 GB of RAM.

To obtain the results reported in the article run "python RUN.py FULL".

The script generates pdf files with graphs depicting the numbers of false 
positive results in a fixed number of candidates. The solid line represents 
results without the use of replication groups, the broken line results with 
two replication groups and the dotted line results with three replication 
groups (as the graphs in the article).

Lines in the standard output, which contain "AUCSOUT", provide AUC scores for
the given combination of interaction scoring method and the dataset. First 
the score without the use of replication groups is reported, afterwards the 
score with two replication groups, and the last score is the AUC obtained with
three replication groups.

