HyperND

This repo contains the implementation for the algorithm HyperND from the paper:

Nonlinear Feature Diffusion on Hypergraphs
By Konstantin Prokopchik, Austin R. Benson and Francesco Tudisco \

To be presented at ICML 2022.

Baselines

To install required julia packages run julia packages.jl or include(packages.jl) (if you are in a julia terminal).

We have gathered 5 baselines, each of their realizations is taken from a github page.

1. APPNP
2. HGNN
3. HyperGCN
4. SCE
5. SGC

We have wrapped them into packages for convenience, the guide to installation inside competitors/competitors_setups directory.

There are 3 experiments:

1. HyperGCN experiment is in the old format in the root folder, that extensively compares HyperGCN with our algorithm.
   Execute julia cross_val_datasets_HOLS_ft.jl or include("cross_val_datasets_HOLS_ft.jl") to reproduce.
2. Time experiment compares times of our algorithm and baselines.
   Execute julia competitors/scripts/time.jl or include("competitors/scripts/time.jl") to reproduce.
3. Main experiment does a CV for our algorithm and compares the results with all the baselines on the same input data across multiple runs. Execute julia competitors/scripts/main.jl or include("competitors/scripts/main.jl") to reproduce.

The datasets for experiments are inside the data folder. Results are stored in competitors/results. All the additional information is inside the scripts.