This repository contains the implementation of Sheaf-enhanced Hypergraph Neural Networks

We recommend using Conda to manage the environment.

1. Create and activate a new environment:

   conda create -n sheaf_env python=3.10
   conda activate sheaf_env

2. Install dependencies: Ensure you have the requirements.txt file in the root directory.

   pip install -r requirements.txt

   Note: Please ensure the directory structure is maintained for local imports (e.g., models, layers, helper, sheaf_builder.py, sheafRHNN_conv.py).

The datasets for link prediction tasks are pre-packaged in this repository under the data/ folder:

• FB-AUTO: data/FB-AUTO
• M-FB15K: data/M-FB15K

For node classification tasks (e.g., Cora, PubMed, Citeseer), we utilize the preprocessing steps from the AllSet framework.

Setup Instructions: To generate datasets, please follow the instructions from the AllSet Repository. Once generated, place the data folder in the root directory of this project (e.g., ./data/).

This repository provides two main entry scripts.

Run the model for link prediction.

python run_inference.py \
  -data data/FB-AUTO \
  -score_func conve \
  -opn corr \
  -gpu 0 \
  -epoch 5 \
  -batch 256

Run the model for node classification.

python run_classification.py \
  --dname pubmed \
  --method sheafHyperGNNDiag \
  --cuda 0 \
  --epochs 10 \
  --heads 2 \
  --MLP_hidden 32

We provide Shell scripts to reproduce the experiments regarding dimension variations and noise robustness (SNR tests).

Prerequisite: Ensure the scripts are executable:

chmod +x *.sh

1. Dimension Variation Test (Classification) Runs the classification model across different hidden dimensions (2, 4, ..., 128).

   ./run_dim_test.sh

2. Noise Robustness Test (Classification) Evaluates the model under Gaussian and Rayleigh noise with varying SNR (-6dB to 10dB).

   ./run_class_noise.sh

3. Noise Robustness Test (Inference) Evaluates the link prediction model under different noise conditions.

   ./run_infer_noise.sh

If you find this code useful for your research, please consider citing our work.