GStarX: Explaining Graph Neural Networks with Structure-Aware Cooperative Games

Authors: Shichang Zhang, Yozen Liu, Neil Shah, Yizhou Sun

NeurIPS 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We demonstrate that GStar X produces qualitatively more intuitive explanations, and quantitatively improves explanation fidelity over strong baselines on chemical graph property prediction and text graph sentiment classification. We conduct experiments on datasets from different domains including synthetic graphs, chemical graphs, and text graphs. Quantitative studies. We report averaged test set H-Fidelity in Table 1. Qualitative studies. We visualize the explanations of graphs in Graph SST2 in Figure 2 and compare them qualitatively. Ablation study and analysis.
Researcher Affiliation Collaboration Shichang Zhang1 Yozen Liu2 Neil Shah2 Yizhou Sun1 1University of California, Los Angeles 2Snap Inc. 1{shichang, yzsun}@cs.ucla.edu 2{yliu2, nshah}@snap.com
Pseudocode Yes Algorithm 1 GStar X: Graph Structure-Aware Explanation. Algorithm 2 The Compute-HN Function.
Open Source Code Yes Code available at https://github.com/ShichangZh/GStarX
Open Datasets Yes We conduct experiments on datasets from different domains including synthetic graphs, chemical graphs, and text graphs. MUTAG [8], BACE and BBBP [39] contain chemical molecule graphs. Graph SST2 and Twitter [43] contain graphs constructed from text. BA2Motifs [25] contains graphs with a Barabasi-Albert (BA) base graph.
Dataset Splits Yes We use a 10-fold cross-validation setup for all datasets.
Hardware Specification No The paper does not provide specific details about the hardware used to run the experiments (e.g., GPU models, CPU types, or cloud instance specifications).
Software Dependencies No The paper mentions software like PyTorch Geometric, BERT embeddings, and Biaffine parser, but does not provide specific version numbers for these or other dependencies required for reproducibility.
Experiment Setup Yes All models are trained to convergence with hyperparameters and performance shown in Appendix A.2. Our models were trained for 300 epochs for all datasets using Adam optimizer with an initial learning rate of 0.01 and weight decay of 0.0001.