Fast Feature Selection for Learning to Rank

This is the Python 2.7 code used in "Fast feature selection for Learning to Rank" (2016) by A. Gigli, C. Lucchese, R. Perego, F. Nardini.

You can clone it ad run in bash through the following command

• 0_prepare_data.sh <YOUR_DIRECTORY> <FEATURES_NUMBER>

• 1_run_fast_feature_selection.sh <YOUR_DIRECTORY>

• 2_run_competitors.sh <YOUR_DIRECTORY>

• 3_effectiveness_performance.sh <YOUR_DIRECTORY> <FEATURES_FILE>

where

• <YOUR_DIRECTORY> is the path of the directory where you have cloned this repository.
• <FEATURES_NUMBER> is the number of feature of the dataset.
• <FEATURES_FILE> is the filename of the file containing the selected features. In the paper we compare different Feature Selection Algorithms (FSAs): GAS, NGAS, XGAS and HCAS. In order to make the performance measurement more flexible 1_run_fast_feature_selection.sh <YOU_DIRECTORY> generates 4 files: gas_selection_test.txt, ngas_selection_test.txt, xgas_selection_test.txt, hcas_selection_test.txt in the folder /output. Each file represents the selection obtained from a FSA and contains 7 feature selections (one for each features selection subset dimension, ie 5%, 10%, 20%, 30%, 40%, 50%, 75% ).

The code demo is ready to be tested on sample data extracted from the *istella LETOR dataset. The dataset contains 220 features in SVM format, so please set <FEATURES_NUMBER> equal to 220 when launching 0_prepare_data.sh <YOUR_DIRECTORY> <FEATURES_NUMBER>

HINT: If you simply want to run LambdaMART on the whole features set you can create a one-row file containing a sequence of integers from 1 to <FEATURES_NUMBER> separated by white spaces and then run from your /test_feature_selection_performance local directory

python test_performance_lambdamart.py <DATA_DIRECTORY> <FEATURES_FILE>

where

• <DATA_DIRECTORY> is the directory where you have saved train, validation and test data in the appropriate format
• <FEATURES_FILE> is the one-row file containg the sequence of integers