Adaptive Channel Mixing (ACM) GNN Framework

Revisiting Heterophily for Graph Neural Networks (Sitao Luan et al., NeurIPS 2022): https://arxiv.org/abs/2210.07606

Updated manuscript with additional results on large-scale datasets (Penn94, arXiv-year, genius, twitch-gamers). See Table 4 in Appendix A.3.

Homophily metrics	t-SNE visualization of GCN and ACM-GCN layers trained on Squirrel

Repository Overview

We provide the PyTorch implementation for ACM-GNN framework here. The repository is organised as follows:

|-- ACM-PyTorch # experiments on 10 small-scale datasets
    |-- experiment/ # experiment bash script 
    |-- models/ # model definition
    |-- splits/ # split files for datasets
    |-- arg_parser.py  # the argument parser code for training/hyperparameter searching script
    |-- hyperparameter_searching.py # the hyperparameter searching script
    |-- logger.py # the logger code
    |-- train.py # the model trainig script
    |-- utils.py # data process and others
|-- ACM-Geometric # experiments on the large-scale and small-scale datasets based on the data provided by LINKX
    |-- sh/ # experiment bash script
    |-- large_scale_data/ # dataset folder
    |-- train.py # the model trainig script
    |-- parse.py  # set hyper-parameters
|-- data/ # 3 old datasets, including cora, citeseer, and pubmed
|-- new-data/ # 6 new datasets, including texas, wisconsin, cornell, actor, squirrel, and chameleon
|-- synthetic-experiments # generate synthetic features and graphs with different homophily levels and train baseline models
    |-- baseline_models/ # models and layers of baseline models
    |-- feature_generation.py  # generate node features with base datasets
    |-- graph_generation.py  # generate graphs with different homophily levels
    |-- train.py # the training code
    |-- utils.py # data process and others
    |-- hyperparameter_searching.py # searching the hyperparameters of the baseline model
    |-- logger.py # logger
    |-- homophily.py # different homophily metrics
|-- plots # all experimental plots and visualizations in our paper

Dependencies

The script has been tested running under Python 3.7.4, with the following packages installed (along with their dependencies):

dgl-cpu==0.4.3.post2
dgl-gpu==0.4.3.post2
ogb==1.3.1
numpy==1.19.2
scipy==1.4.1
networkx==2.5
torch==1.5.0
torch-cluster==1.5.7
torch-geometric==1.6.3
torch-scatter==2.0.5
torch-sparse==0.6.6
torch-spline-conv==1.2.0

In addition, CUDA 10.2 has been used.

pip install -r requirements.txt

Running Experiments (ACM-PyTorch)

Training or hyperparameter searching

# training with default hyperparameters (e.g. ACM-GCN+ on Texas)
python train.py --model acmgcnp --dataset_name texas

# training with user defined hyperparameters
python train.py --model acmgcnp --dataset_name texas --lr 0.06 --weight_decay 0.0006 --dropout 0.6

# hyperparameter searching (learning rate, weight_decay & dropout)
python hyperparameter_searching.py --model acmgcnp --dataset_name texas

The training/hyperparameter seacrhing logs are saved into the logs/ folder located at <your-working-directory>/Adaptive-Channel-Mixing-GNN/ACM-PyTorch/logs.

Running Experiments (ACM-Geometric)

Create a results/ folder at <your-working-directory>/Adaptive-Channel-Mixing-GNN/ACM-Geometric. Experimental results (.csv) would be saved here.

Download the data/ folder from the repository of LINKX to <your-working-directory>/Adaptive-Channel-Mixing-GNN/ACM-Geometric/large_scale_data.

# training with default hyperparameters (e.g. ACM-GCN+ on twitch-gamer)
python train.py --model acmgcnp --dataset_name "twitch-gamer"

# training with user defined hyperparameters
python train.py --model acmgcnp --dataset_name "twitch-gamer" --lr 0.002 --weight_decay 0.0006 --dropout 0.6

# hyperparameter searching
bash sh/run_all_settings.sh

Synthetic Benchmark Experiments

1. Generate features & graphs

# generate features
python feature_generation.py --base_dataset cora

# generate (random or regular) graphs
python graph_generation.py --graph_type random --degree_intra 2 --num_graph 10

Generated features are saved into the synthetic_graphs/features/ folder located at <your-working-directory>/Adaptive-Channel-Mixing-GNN/synthetic-experiment/synthetic_graphs/.

Generated graphs are saved into the synthetic_graphs/<random|regular>/ folder located at <your-working-directory>/Adaptive-Channel-Mixing-GNN/synthetic-experiment/synthetic_graphs/<random|regular>.

2. Training or hyperparameter searching

# training with user defined hyperparameters
python train.py --model_type acmgcn --base_dataset cora --graph_type random --edge_homo 0.1 --degree_intra 2 --num_graph 10 --lr 0.05 --weight_decay 0.0005 --dropout 0.5

# hyperparameter searching (weight_decay & dropout)
python hyperparameter_searching.py --model_type acmgcn --graph_type random --base_dataset cora --edge_homo 0.1 --edge_homo 0.1 --degree_intra 2 --num_graph 10

Graph generation and training logs are saved into the logs/ folder located at <your-working-directory>/Adaptive-Channel-Mixing-GNN/synthetic-experiment/logs.

Papers With Code Leaderboard

See the leaderboards for node classification on heterophilic graphs here. Feel free to add your results or create new benchmark datasets.

Attribution

Parts of the code are based on

GCN
LINKX

Reference

If you make advantage of the ACM framework in your research, please cite the following in your manuscript:

@article{luan2022revisiting,
  title={Revisiting heterophily for graph neural networks},
  author={Luan, Sitao and Hua, Chenqing and Lu, Qincheng and Zhu, Jiaqi and Zhao, Mingde and Zhang, Shuyuan and Chang, Xiao-Wen and Precup, Doina},
  journal={Advances in neural information processing systems},
  volume={35},
  pages={1362--1375},
  year={2022}
}

License

MIT

SitaoLuan / ACM-GNN