Nonnegative binary matrix factorization with majorization-minimization

This repository contains the code for reproducing the experiments in our paper entitled A majorization-minimization algorithm for nonnegative binary matrix factorization, published in the IEEE Signal Processing Letters in 2022.

After cloning or downloading this repository, you will have the .rda data for several datasets.

Requirements

In order to run the experiments (which uses the R package `logisticPCA), you need to install R. The detailed instructions are available here, but minimally:

sudo apt install dirmngr gnupg apt-transport-https ca-certificates software-properties-common
sudo apt-key adv --keyserver keyserver.ubuntu.com --recv-keys E298A3A825C0D65DFD57CBB651716619E084DAB9
sudo add-apt-repository 'deb https://cloud.r-project.org/bin/linux/ubuntu focal-cran40/'
sudo apt install r-base

Then run R (e.g., in command line, simply type R), and install the corresponding pacakges:

install.packages('logisticPCA')
install.packages('rARPACK')

Finally, you also need to install the Python package rpy2 for interfacing both.

Reproducing the experiments

Now that you're all set, simply run the following scripts:

• main_script.py trains the models (NBMF-EM, NBMF-MM, and logPCA) while performing validation (minimize perplexity on the validation set), to tune the hyperparameters. It also predict the test data with various random initializations.

• display_results.py displays the plots from the paper (validation score for NBMF-MM, test performance for all methods, and estimated H matrix on the lastFM dataset).

Reference

If you use any of this code for your research, please cite our paper:

@article{Magron2022nbmf,  
  author={P. Magron and C. F{\'e}votte},  
  title={A majorization-minimization algorithm for nonnegative binary matrix factorization},  
  journal={IEEE ISignal Processing Letters},  
  year={2022}
}