GVEX

This repository contains the source code for our paper: View-based Explanations for Graph Neural Networks, SIGMOD 2024, by Tingyang Chen, Dazhuo Qiu, Yinghui Wu, Arijit Khan, Xiangyu Ke, Yunjun Gao

Requirements

---

See requirements.txt

Datasets

---

We use the following datasets in our experiments:

• Mutagenicity
• ENZYMES
• REDDIT-BIANRY
• MALNET-TINY
• PCQM4Mv2

---

Structure

---

• checkpoints: store the trained model.
• config: the parameters of the algorithm and model.
• datasets: datasets used in the experiments.
• approximate_algorithm.py and streaming_algorithm.py: the GVEX algorithms.
• train_gnn.py: train the model.
• utils.py: some help functions.
• visualization.py: visualize the explanations.

---

Usage

---

(For SIGMOD'24 ARI: Step 4 is enough.)

1. Download datasets

2. Configure the trainning para meters and run train_gnn.py to train the model(stored in checkpoints):

    learning_rate: 0.001
    weight_decay: 5e-4
    milestones: None
    gamma: None
    batch_size: 32
    num_epochs: 2000 
    num_early_stop: 0
    gnn_latent_dim:
      - 128 # 128
      - 128
      - 128
    gnn_dropout: 0.0
    add_self_loop: True
    gcn_adj_normalization: True
    gnn_emb_normalization: False
    graph_classification: True
    node_classification: False
    gnn_nonlinear: 'relu'
    readout: 'max'
    fc_latent_dim: [ ]
    fc_dropout: 0.0
    fc_nonlinear: 'relu'
    concate: False
   

   

3. Config the algorithm parameters in config folder:

   dataset_root: 'datasets'
   dataset_name: 'Mutagenicity'
   random_split_flag: True
   data_split_ratio: [0.8, 0.1, 0.1]
   seed: 2
   data_explain_cutoff: -1
   budget: 100
   threshold: 0.08
   gamma: 1
   radium: 0.005
   k: 5
   bounds: [0, 100, 0, 0]
   num_classes: 2
   
   

4. Run approximate_algorithm.py or streaming_algorithm to generate the explanations:

   

   

---

Figures