Explaining Graph Neural Networks via Structure-aware Interaction Index

Authors: Ngoc Bui, Hieu Trung Nguyen, Viet Anh Nguyen, Rex Ying

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on various graph datasets and models demonstrate that our method consistently provides superior subgraph explanations compared to state-of-the-art methods.
Researcher Affiliation Collaboration 1Yale University 2Vin AI Research 3The Chinese University of Hong Kong.
Pseudocode Yes Algorithm 1 Permutation-based sampling algorithm for the k-order Myerson-Taylor index. Algorithm 2 Value of an interaction-restricted function (f|E(T)).
Open Source Code Yes Our implementation is available at: https://github. com/ngocbh/MAGE/
Open Datasets Yes We use ten datasets commonly used in the graph explainability literature, including synthetic data, biological, text, and image data. For synthetic datasets, we use Ba-2Motifs (Luo et al., 2020), BA-House Grid (Amara et al., 2023), and SPMotif (Wu et al., 2022) for classification tasks
Dataset Splits Yes We split the dataset into training, validation, and test subsets with respective ratios of 0.8, 0.1, and 0.1.
Hardware Specification No The paper does not provide specific details about the hardware (e.g., GPU/CPU models, memory specifications) used to run the experiments.
Software Dependencies Yes One strategy to solve (3) is by linear relaxations and then using off-the-shelf MILP solvers such as MOSEK (Ap S, 2019) or GUROBI (Gurobi Optimization, LLC, 2023).
Experiment Setup Yes Regarding hyperparameter settings, we set the number of explanatory nodes M and components m according to the ground truth explanations for all the baselines if they are available. ... The number of permutations used to compute the Myerson-Taylor index is set to 200, and we use MOSEK (Ap S, 2019) with default parameters for the motif search.