GNNExplainer: Generating Explanations for Graph Neural Networks

Authors: Zhitao Ying, Dylan Bourgeois, Jiaxuan You, Marinka Zitnik, Jure Leskovec

NeurIPS 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments on synthetic and real-world graphs show that our approach can identify important graph structures as well as node features, and outperforms alternative baseline approaches by up to 43.0% in explanation accuracy.
Researcher Affiliation Collaboration Rex Ying Dylan Bourgeois , Jiaxuan You Marinka Zitnik Jure Leskovec Department of Computer Science, Stanford University Robust.AI {rexying, dtsbourg, jiaxuan, marinka, jure}@cs.stanford.edu
Pseudocode No The paper describes algorithmic steps and formulations but does not include a formally labeled "Pseudocode" or "Algorithm" block.
Open Source Code Yes Code and datasets are available at https://github.com/Rex Ying/gnn-model-explainer
Open Datasets Yes We construct four kinds of node classification datasets (Table 1). (1) In BASHAPES, we start with a base Barab asi-Albert (BA) graph on 300 nodes and a set of 80 five-node house -structured network motifs, which are attached to randomly selected nodes of the base graph. (2) BA-COMMUNITY dataset is a union of two BA-SHAPES graphs. (3) In TREE-CYCLES, we start with a base 8-level balanced binary tree and 80 six-node cycle motifs. (4) TREE-GRID is the same as TREE-CYCLES except that 3-by-3 grid motifs are attached to the base tree graph. We consider two graph classification datasets: (1) MUTAG is a dataset of 4,337 molecule graphs labeled according to their mutagenic effect on the Gram-negative bacterium S. typhimurium [10]. (2) REDDIT-BINARY is a dataset of 2,000 graphs... [37]. Code and datasets are available at https://github.com/Rex Ying/gnn-model-explainer
Dataset Splits No The paper mentions training models and evaluating results, but does not explicitly describe validation data splits or how validation was performed to tune hyperparameters.
Hardware Specification No The paper does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions
Experiment Setup Yes Hyperparameters KM, KF control the size of subgraph and feature explanations respectively, which is informed by prior knowledge about the dataset. For synthetic datasets, we set KM to be the size of ground truth. On real-world datasets, we set KM = 10. We set KF = 5 for all datasets. We further fix our weight regularization hyperparameters across all node and graph classification experiments.