This directory contains information to perform experiments using the best features selected by the benchmark participants:

- NIPS2003-Dec1-results.xls and NIPS2003-Dec8-results.xls contain the performances of the leading challenge participants. The entries with few features and good scores are highlighted. We selected in the end the entries that satisfied the following criteria: be consistently good accross all data sets (only the Neal small NN satisfed this criterion in both Dec 1 abd 8), or be good for a particular data set. We maximized diversity by allowing a maximum of one feature set by participant for each dataset. We selected feature sets of size below 10% of the total number of features, except if there were no such subsets in the top ranking entries, in which case we went slightly above 10%.

- best_feat.m: These are the directory names of the feature set that were retained (five for each dataset).

- Each of the 50 files xxx.feat contains the indices of selected features. The file name uses the following code:
date_filename_dataset
where
date: dec1 or dec8. dec1 results did not use the knowledge of the validation set. dec8 results did.
filename: identifier of the file where the results were stored on the server.
dataset: name of the dataset.
e.g. dec8_caBnuxYN_dorothea.feat is the dec8 entry stored in caBnuxYN for the dorothea dataset.