| Environment Setup | Quick Start | Paper | Video | Datasets
Before you begin, please make sure that you have Anaconda or Miniconda installed on your system.
# Create and activate a new Conda environment named 'GIP'
conda create -n GIP
conda activate GIP
Download the dependencies from the requirements.txt:
# Install dependencies from files
pip install -r requirements.txt
The Architecture of GIP is shown as follows:
We have provided scripts with hyper-parameter settings to get the experimental results, and you can obtain the experimental results by running the parameters you want:
python -m train --bmname=<dataset> --assign-ratio=<compression_ratio> --hidden-dim=<hiddden_dim> --output-dim=<output_dim> --num-classes=<num_of_classes> --method=soft-assign --epochs=<epochs> --max-nodes=<max_node> --ker-hidden-graphs=<num_of_prototypes> --ker-size-hidden-graphs=<num_of_nodes_in_prototype> --ker-max-step=<max_walk_steps>
The experimental parameter settings for the results in the paper are shown below：
For example:
# ENZYMES
python -m train --bmname=ENZYMES --assign-ratio=0.05 --hidden-dim=24 --output-dim=8 --cuda=0 --num-classes=6 --method=soft-assign --epochs=500 --max-nodes=200 --ker-hidden-graphs=24 --ker-size-hidden-graphs=10
# DD
python -m train --bmname=DD --assign-ratio=0.1 --hidden-dim=64 --output-dim=64 --num-classes=2 --method=soft-assign --epochs=300 --max-node=300 --ker-hidden-graphs=6 --ker-size-hidden-graphs=8 --ker-max-step=3 --feature=label
# PROTEINS
python -m train --bmname=PROTEINS --assign-ratio=0.1 --hidden-dim=16 --output-dim=16 --num-classes=2 --method=soft-assign --epochs=500 --max-nodes=300 --ker-hidden-graphs=8 --ker-size-hidden-graphs=10 --ker-max-step=3
# COLLAB
python -m train --bmname=COLLAB --assign-ratio=0.1 --hidden-dim=16 --output-dim=3 --num-classes=3 --method=soft-assign --epochs=500 --max-nodes=300 --ker-hidden-graphs=15 --ker-size-hidden-graphs=10 --ker-max-step=3
# GraphCycle
python -m train --bmname=GraphCycle --assign-ratio=0.05 --hidden-dim=5 --output-dim=5 --num-classes=2 --method=soft-assign --ker-hidden-graphs=8 --ker-size-hidden-graphs=8 --ker-max-step=3 --feature=deg-num
# GraphFive
python -m train --bmname=FiveClass --assign-ratio=0.05 --hidden-dim=8 --output-dim=5 --num-classes=5 --method=soft-assign --epochs=500 --max-nodes=300 --ker-hidden-graphs=20 --ker-size-hidden-graphs=10 --ker-max-step=3
Please download the dataset from https://chrsmrrs.github.io/datasets/docs/home/
| Dataset | Graphs | Avg.nodes | Avg.edges | Graph classes | Domain |
|---|---|---|---|---|---|
| ENZYMES | 600 | 32.63 | 62.14 | 6 | proteins |
| D&D | 1178 | 284.32 | 715.66 | 2 | proteins |
| PROTEINS | 1113 | 39.06 | 72.82 | 2 | proteins |
| MUTAG | 188 | 17.93 | 19.79 | 2 | molecular |
| COLLAB | 5000 | 74.49 | 2457.78 | 3 | molecular |
| GraphCycle | 2000 | 297.70 | 697.18 | 2 | synthetic |
| GraphFive | 5000 | 375.98 | 1561.77 | 5 | synthetic |
🌹Please cite our work if it is helpful:
@article{wang2024unveiling,
title={Unveiling Global Interactive Patterns across Graphs: Towards Interpretable Graph Neural Networks},
author={Wang, Yuwen and Liu, Shunyu and Zheng, Tongya and Chen, Kaixuan and Song, Mingli},
journal={arXiv preprint arXiv:2407.01979},
year={2024}
}