On Explainability of Graph Neural Networks via Subgraph Explorations

Authors: Hao Yuan, Haiyang Yu, Jie Wang, Kang Li, Shuiwang Ji

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our Subgraph X achieves significantly improved explanations, while keeping computations at a reasonable level.
Researcher Affiliation Academia 1Department of Computer Science & Engineering, Texas A&M University, TX, USA 2Department of Electronic Engineering and Information Science, University of Science and Technology of China, Hefei, China 3West China Biomedical Big Data Center, West China Hospital, Chengdu, China.
Pseudocode Yes Finally, we conclude the computation steps of our proposed Subgraph X in Algorithm 1 and 2.
Open Source Code Yes Our code and data are now publicly available in the DIG library (Liu et al., 2021)1. 1https://github.com/divelab/DIG
Open Datasets Yes MUTAG (Debnath et al., 1991) and BBBP (Wu et al., 2018) are molecular datasets for graph classification tasks. ...Graph-SST2 (Yuan et al., 2020c) is sentiment graph dataset...
Dataset Splits No The paper describes the datasets used (MUTAG, BBBP, GRAPH-SST2, BA-2MOTIFS, BA-SHAPE) and general experimental settings, but does not explicitly provide specific training, validation, or test split percentages or sample counts in the main text.
Hardware Specification Yes We conduct our experiments using one Nvidia V100 GPU on an Intel Xeon Gold 6248 CPU.
Software Dependencies Yes Our implementations are based on Python 3.7.6, Py Torch 1.6.0, and Torch-geometric 1.6.3.
Experiment Setup Yes For our proposed Subgraph X and other algorithms with MCTS, the MCTS iteration number M is set to 20. To explore a suitable trade-off between exploration and exploitation, we set the hyperparameter λ in Eq.(3) to 5 for Graph-SST2 (GATs) and BBBP (GCNs) models, and 10 for other models. Since all GNN models contain 3 network layers, we consider 3-hop computational graphs to compute Shapley values for our Subgraph X. For the Monte-Carlo sampling in our Subgraph X, we set the Monte-Carlo sampling steps T to 100 for all datasets.