This repository contains code to accompany the paper "Finding Bipartite Components in Hypergraphs", published in NeurIPS 2021. It provides an implementation of the proposed algorithm based on the new hypergraph diffusion process, as well as the baseline algorithm based on the clique reduction.

Below, you can find instructions for running the code which will reproduce the results reported in the paper.

Feel free to contact me with any questions or comments at peter.macgregor@ed.ac.uk.

The code was written to work with Python 3.6, although other versions of Python 3 should also work. We recommend that you run inside a virtual environment.

To install the dependencies of this project, run

pip install -r requirements.txt

In order to demonstrate our algorithm, you can view the visualisation of the 2-graph constructed at each step by running

python show_visualisation.py

This example was used to create Figure 1 in the paper.

In this section, we give instructions for running the experiments reported in the paper.

We are unfortunately not able to share the data used for the Penn Treebank experiment, and so we give instructions here for how to preprocess this data for use with our code. You will need to have your own access to the Penn Treebank corpus.

Follow the instructions in this repository, passing the --task pos command line option to generate the files train.tsv, test.tsv, and dev.tsv. Copy these three files to the data/nlp/penn-treebank directory.

To run the experiments on real-world data, you should run

python run_experiment.py {experiment_name}

where {experiment_name} is one of 'ptb', 'dblp', 'imdb', or 'wikipedia' to run the Penn Treebank, DBLP, IMDB and Wikipedia experiments respectively.

To run an experiment on a single synthetic hypergraph, run

python run_experiment_synthetic.py {n} {r} {p} {q}

where {n} is the number of vertices in the hypergraph, {r} is the rank of the hypergraph, {p} is the probability of an edge inside a cluster, and {q} is the probability of an edge between clusters. Be careful not to set p or q to be too large. See the main paper for more information about the random hypergraph model. This will construct the hypergraph if needed, and report the performance of the diffusion algorithm and the clique algorithm on the constructed hypergraph.

The full results from our experiments on synthetic hypergraphs are provided in the data/sbm/results directory, along with a Mathematica notebook for viewing them, and plotting the figures shown in the paper.