NewtonNet: Shape-aware Graph Spectral Learning

An official PyTorch implementation of the paper Shape-aware Graph Spectral Learning.

In this repository, we implement the NewtonNet with Shape-aware Regularization.

Contribution

• We are the first to establish a well-defined relationship between graph frequencies and homophily ratios. We empirically and theoretically show that the more homophilous the graph is, the more beneficial the low-frequency is; while the more heterophilous the graph is, the more beneficial the high-frequency is.
• We propose a novel framework NewtonNet using Newton Interpolation with shape-aware regularization that can learn better filter encourages beneficial frequency and discourages harmful frequency, resulting in better node representations.
• Extensive experiments demonstrate the effectiveness of NewtonNet in various settings.

Installation

Create a new virtual environment.

conda create --name NewtonNet python=3.10

Optinal 1: Install via requirements.txt

pip install -r requirements.txt

Optional 2: Install manually.

conda install pytorch torchvision torchaudio pytorch-cuda=11.8 -c pytorch -c nvidia
conda install pyg -c pyg
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-2.1.0+cu118.html
pip install gdown pandas

Running Experiments

To reproduce the results in Table 1, use commands in scripts.sh

python main.py --dataname chameleon --lr 0.05 --temp_lr 0.01 --K 5 --hidden 64 --weight_decay 0.0 --dropout 0.0 --dprate 0.0 --gamma 3 --gamma2 5 --gamma3 0

To search your own optimal hyperparameters, use

python search_multi.py --dataname chameleon
python search_binary.py --dataname genius

To reproduce the results in Fig. 1, run

bash run_train_case.sh