Sparse GEMINI

Article: Sparse and geometry-aware generalisation of the mutual information for joint discriminative clustering and feature selection

Welcome to the Sparse GEMINI repo. This repo contains the official implementation of the sparse gemini algorithm as well as all the commands to reproduce the experiments.

If your use case for Sparse GEMINI is rather for small datasets, you may want to check the numpy implementation in GemClus that provides a ready-to-use API.

Model definition

If you are only interested in using the sparse gemini model, you can simply download the *.py files in the sparse_gemini folder. The packages required to run the model are described in requirements.txt. Once installed, the command

python sparse_gemini/main.py -h

will give details on all parameters that can be used to toy with the model.

Other folders

The remainder of the folders are dedicated to reproducing the experiments of our article:

• utils contains additional scripts for dataset creation, result gathering.
• analysis contains the necessary script to re-obtain the contents of the figures.
• configs contains a set of configuration files for each different experiment.

Other scripts details

There are a couple scripts in the utils folder that are useful for the analysis or the snakemake pipeline during experiments.

• compute_distances
• create_dataset
• extract_common_features
• extract_logreg_feature_history
• extract_mnist_feature_importance
• fetch_mnist
• fetch_openml
• merge_clustering
• merge_clusterings_v2
• merge_selections
• retrieve_optimal_solutions

Redoing experiments

If you are interested in replicating some experiments, please refer to the main file How to redo all experiments.md which lists the step-by-step command lines for all experiments.