GNAM-RFF: Interpretable and Parameter Efficient Graph Neural Additive Models with Random Fourier Features

Project Structure

experiments/
├── models/                  
│   ├── gnamrff.py           
│   ├── gnamrff_graph.py     
│   └── __init__.py
├── utils/                   
│   ├── data_loader.py      
│   ├── trainer.py          
│   └── __init__.py
├── run_node_experiments.sh
├── run_graph_experiments.sh
├── scripts/                 
│   ├── train_node.py           
│   └── train_graph.py                 
├── results/                 
├── requirements.txt        
└── README.md              

Installation

Install dependencies from the requirements file:

pip install -r requirements.txt

For GPU support, first install PyTorch with CUDA from pytorch.org, then install the remaining dependencies.

Quick Start

Node Classification (Cora, Citeseer, Pubmed)

Single dataset:

python scripts/train_node.py --dataset cora --cuda 0

With custom hyperparameters:

python train_node.py \
    --dataset citeseer \
    --cuda 0 \
    --lr 0.05 \
    --weight_decay 5e-4 \
    --M 100 \
    --R 5 \
    --epochs 1000 \
    --save_model

Run all node datasets:

bash scripts/run_node_experiments.sh

Graph Classification (Mutagenicity)

5-fold cross-validation:

python scripts/train_graph.py --dataset mutagenicity --cuda 0 --n_folds 5

With custom hyperparameters:

python scripts/train_graph.py \
    --dataset mutagenicity \
    --cuda 0 \
    --n_folds 5 \
    --lr 0.01 \
    --weight_decay 1e-3 \
    --batch_size 32 \
    --M 100 \
    --R 5 \
    --epochs 400 \
    --save_model

Run all graph datasets:

bash scripts/run_graph_experiments.sh

Citation

If you use this code in your research, please cite:

@article{reddy2025interpretable,
  title={Interpretable and Parameter Efficient Graph Neural Additive Models with Random Fourier Features},
  author={Thummaluru Siddartha Reddy, Vempalli Naga Sai Saketh and Mahesh Chandran},
  booktitle={The Thirty-ninth Annual Conference on Neural Information Processing Systems},
  year={2025},
  url={https://openreview.net/forum?id=yl9LxRL5tj}
}