[Paper Review] GISExplainer: On Explainability of Graph Neural Networks via Game-theoretic Interaction Subgraphs

Xingping Xian, Jing Liu+5 more|arXiv (Cornell University)|Sep 24, 2024

Scientific Computing and Data ManagementDecision Sciences3 citations

TL;DR

This paper proposes GraphGI, a game-theoretic explanation method for Graph Neural Networks (GNNs) that identifies subgraphs with the highest interaction strength among edges, using Shapley-based game-theoretic interaction values to guide iterative edge selection. The method achieves superior fidelity and sparsity compared to state-of-the-art baselines by modeling edge-level interactions rather than relying on node-centric or independent feature attribution.

ABSTRACT

Explainability is crucial for the application of black-box Graph Neural Networks (GNNs) in critical fields such as healthcare, finance, cybersecurity, and more. Various feature attribution methods, especially the perturbation-based methods, have been proposed to indicate how much each node/edge contributes to the model predictions. However, these methods fail to generate connected explanatory subgraphs that consider the causal interaction between edges within different coalition scales, which will result in unfaithful explanations. In our study, we propose GISExplainer, a novel game-theoretic interaction based explanation method that uncovers what the underlying GNNs have learned for node classification by discovering human-interpretable causal explanatory subgraphs. First, GISExplainer defines a causal attribution mechanism that considers the game-theoretic interaction of multi-granularity coalitions in candidate explanatory subgraph to quantify the causal effect of an edge on the prediction. Second, GISExplainer assumes that the coalitions with negative effects on the predictions are also significant for model interpretation, and the contribution of the computation graph stems from the combined influence of both positive and negative interactions within the coalitions. Then, GISExplainer regards the explanation task as a sequential decision process, in which a salient edges is successively selected and connected to the previously selected subgraph based on its causal effect to form an explanatory subgraph, ultimately striving for better explanations. Additionally, an efficiency optimization scheme is proposed for the causal attribution mechanism through coalition sampling. Extensive experiments demonstrate that GISExplainer achieves better performance than state-of-the-art approaches w.r.t. two quantitative metrics: Fidelity and Sparsity.

Motivation & Objective

• To address the limitation of existing GNN explanation methods that treat features as independent, ignoring their interactions.
• To improve model interpretability by identifying subgraphs that capture the strongest cooperative interactions among edges.
• To develop an efficient computational framework for game-theoretic interaction values that scales to real-world graphs.
• To enhance fidelity and sparsity in explanations by modeling edge-level interactions within the GNN’s receptive field.

Proposed method

• GraphGI formulates GNN explanation as a cooperative game where edges are players, and their collective contribution is measured via game-theoretic interaction values.
• It incrementally builds an explanation subgraph by selecting edges that maximize the interaction strength of the current coalition.
• The method uses Shapley values and their extensions to compute interaction strengths, ensuring fair attribution of cooperative effects.
• To reduce computational cost, it applies Monte Carlo sampling for approximating Shapley values and interaction values.
• The explanation subgraph is constrained to the L-hop neighborhood of the prediction target, aligning with GNN message-passing mechanisms.
• The algorithm prioritizes sparse yet highly interactive edge sets, favoring single directed edges over node-connected subgraphs.

Experimental results

Research questions

• RQ1 Can edge-level interactions in GNNs be effectively modeled to improve explanation fidelity?
• RQ2 How does modeling cooperative interactions among edges compare to independent feature attribution in GNN explanation?
• RQ3 To what extent can game-theoretic interaction values enhance the sparsity and interpretability of GNN explanations?
• RQ4 Can efficient approximation techniques maintain high accuracy while reducing computational cost in interaction-based explanation?

Key findings

• GraphGI achieves the highest fidelity and sparsity on both synthetic and real-world datasets, outperforming SubgraphX, PGExplainer, and other baselines.
• On the BA-shapes and Tree-cycles datasets, GraphGI achieved the best fidelity and sparsity under the GIN model, demonstrating robustness across graph structures.
• In real-world datasets like Cora and CiteSeer, GraphGI maintained high performance even at high sparsity levels, indicating effective identification of critical interaction subgraphs.
• The method produced sparser explanations than SubgraphX and PGExplainer due to its edge-centric, interaction-driven selection strategy.
• The use of Monte Carlo sampling significantly reduced computation time without sacrificing explanation quality, enabling scalability to larger graphs.