Hypergraph Diffusions

Step 1: Create a dataset

Inside the data directory, run the following command

python hypergraphs2hmetis.py -vv

Step 2: Run the animation

From the top level you can run this command

python animate_diffusion.py -g data/fauci_email.hmetis -l data/fauci_email.label -f infinity -v

Manifold learning experiments

No setup action necessary, see notebook for demo.

To generate the figures in the ICML paper, run

python ICML2024_Figure_Generation.py

Fast C++ implementation

For using the fast C++ implementation you will first need to install the Eigen library

sudo apt install libeigen3-dev

Or download it through here. Make sure that the Eigen folder is available in /usr/include.

Install pybind11

python3 -m pip install pybind11

Compile the python wrapper

python diffusion_wrapper_setup.py build_ext --inplace

Opening a python terminal you can run a hypergraph diffusion as follows

import reading
import diffusion


n, m, node_weights, hypergraph, weights, center_id, hypergraph_node_weights = reading.read_hypergraph('data/Paper_datasets/zoo.hmetis')
# Turn hypergraph into list of lists
hypergraph = [list(e) for e in hypergraph]
label_names, labels = reading.read_labels('data/Paper_datasets/zoo.label')
gs = diffusion.GraphSolver(n, m, node_weights, hypergraph, len(label_names), labels, 0)
gs.run_diffusions("zoo", 1, 100, 1, 0.1, 20, 5, 50, 0)

Alternatively, you can compile and use the C++ driver

g++ -std=c++20 -Wall -Wextra -O3 -o fast_alg diffusion.cpp fast_alg.cpp

An example call is

/fast_alg --graph_filename data/Paper_datasets/covertype45.hmetis --label_filename data/Paper_datasets/covertype45.label -T 3000 --lambda 1 -h 0.4 --minimum_revealed 20 --step 20 --maximum_revealed 200 -r 5 -e 40 --schedule 0 -v 2> covertype45_all.txt