How Powerful are Spectral Graph Neural Networks

This repository is the official implementation of the model in the following paper:

Xiyuan Wang, Muhan Zhang: How Powerful are Spectral Graph Neural Networks. ICML 2022

@article{JacobiConv,
  author    = {Xiyuan Wang and
               Muhan Zhang},
  title     = {How Powerful are Spectral Graph Neural Networks},
  journal   = {ICML},
  year      = {2022}
}

Requirements

Tested combination: Python 3.9.6 + PyTorch 1.9.0 + PyTorch_Geometric 2.0.3 + PyTorch Sparse 0.6.12

Other required python libraries include: numpy, scikit-learn, optuna, seaborn etc.

Reproduce Our Results

Image Filter Tasks

To reproduce results of JacobiConv on image datasets:

python ImgFilter.py --test --repeat 1 --dataset $dataset --fixalpha

where $dataset is selected from low, high, rejection, band, and comb.

To reproduce results of linear GNN with other bases:

python ImgFilter.py --test --$basis --repeat 1 --dataset $dataset --fixalpha

where $basis is selected from cheby, power, and bern.

We use optuna to select hyperparameters.

python ImgFilter.py --optruns 100  --dataset $dataset  --path $dir --name $dataset

The record file of optuna will be put in directory $dir.

Real-World Tasks

To reproduce results of JacobiConv on real-world datasets:

python RealWorld.py --test --repeat 10 --dataset $dataset --split dense

where $dataset is selected from pubmed, computers, squirrel, photo, chameleon, film, cora, citeseer, texas, cornell.

To reproduce results of linear GNN with other bases:

python RealWorld.py --test --$basis  --fixalpha --repeat 10 --dataset $dataset --split dense

where $basis is selected from cheby, power, and bern.

To reproduce other ablation studies:

Unifilter

python RealWorld.py --test --repeat 10 --dataset $dataset --split dense --sole

No-PCD

python RealWorld.py --test --repeat 10 --dataset $dataset --split dense --fixalpha

NL-RES

python RealWorld.py --test --repeat 10 --dataset $dataset --split dense --resmultilayer

python RealWorld.py --test --repeat 10 --dataset $dataset --split dense --multilayer

To select hyperparameters: