Multi-level Disentanglement Graph Neural Network (MD-GNN)

This is a PyTorch implementation of the MD-GNN, and the code includes the following modules:

• Datasets (Cora, Citeseer, Pubmed, Synthetic, and ZINC)

• Training paradigm for node classification, graph classification, and graph regression tasks

• Visualization

• Evaluation metrics

Main Requirements

• dgl==0.4.3.post2
• networkx==2.4
• numpy==1.18.1
• ogb==1.1.1
• scikit-learn==0.22.2.post1
• scipy==1.4.1
• torch==1.5.0

Description

• train.py

  • main() -- Train a new model for node classification task on the Cora, Citeseer, and Pubmed datasets
  • evaluate() -- Test the learned model for node classification task on the Cora, Citeseer, and Pubmed datasets
  • main_synthetic() -- Train a new model for graph classification task on the Synthetic dataset
  • evaluate_synthetic() -- Test the learned model for graph classification task on the Synthetic dataset
  • main_zinc() -- Train a new model for graph regression task on the ZINC datasets
  • evaluate_zinc() -- Test the learned model for graph regression task on the ZINC datasets

• dataset.py

  • load_data() -- Load data of selected dataset

• MDGNN.py

  • MDGNN() -- model and loss

• utils.py

  • evaluate_att() -- Evaluate attribute-level disentanglement with the visualization of relation-related attributes
  • evaluate_corr() -- Evaluate node-level disentanglement with the correlation analysis of latent features
  • evaluate_graph() -- Evaluate graph-level disentanglement with the visualization of disentangled relation graphs

Running the code

1. Install the required dependency packages and unzip files in the data folder.

2. We use DGL to implement all the GNN models on three citation datasets (Cora, Citeseer, and Pubmed). In order to evaluate the model with different splitting strategy (fewer and harder label rates), you need to replace the following file with the citation_graph.py provided.

dgl/data/citation_graph.py

3. To get the results on a specific dataset, run with proper hyperparameters

python train.py --dataset data_name

where the data_name is one of the five datasets (cora, citeseer, pubmed, synthetic, and zinc). The model as well as the training log will be saved to the corresponding dir in ./log for evaluation.

4. The evaluation the performance of three-level disentanglement performance, run

python utils.py

License

MD-GNN is released under the MIT license.