[Paper Review] On Explainability of Graph Neural Networks via Subgraph Explorations

Hao Yuan, Haiyang Yu+3 more|arXiv (Cornell University)|Feb 9, 2021

Explainable Artificial Intelligence (XAI)Computer Science49 references89 citations

TL;DR

SubgraphX explains GNN predictions by identifying important connected subgraphs using Monte Carlo Tree Search and Shapley-value-based scoring with efficient approximations, improving explanations with reasonable compute.

ABSTRACT

We consider the problem of explaining the predictions of graph neural networks (GNNs), which otherwise are considered as black boxes. Existing methods invariably focus on explaining the importance of graph nodes or edges but ignore the substructures of graphs, which are more intuitive and human-intelligible. In this work, we propose a novel method, known as SubgraphX, to explain GNNs by identifying important subgraphs. Given a trained GNN model and an input graph, our SubgraphX explains its predictions by efficiently exploring different subgraphs with Monte Carlo tree search. To make the tree search more effective, we propose to use Shapley values as a measure of subgraph importance, which can also capture the interactions among different subgraphs. To expedite computations, we propose efficient approximation schemes to compute Shapley values for graph data. Our work represents the first attempt to explain GNNs via identifying subgraphs explicitly and directly. Experimental results show that our SubgraphX achieves significantly improved explanations, while keeping computations at a reasonable level.

Motivation & Objective

• Motivate the need for subgraph-level explanations in GNNs beyond node/edge explanations.
• Propose SubgraphX to identify important connected subgraphs that explain GNN predictions.
• Develop an efficient Shapley-value-based scoring and Monte Carlo Tree Search framework for subgraph exploration.
• Demonstrate effectiveness and efficiency across graph classification and node classification tasks.

Proposed method

• Formulate explanation as finding the most important connected subgraph under a size cap using Score(f, G, Gi).
• Use Monte Carlo Tree Search to explore subgraphs, with pruning-based actions and a UCB-inspired selection rule.
• Adopt Shapley values as the subgraph importance score to capture interactions among graph structures.
• Approximate Shapley values by restricting interactions to L-hop neighborhoods and by Monte Carlo sampling over coalitions.
• Compute marginalized contributions via zero-padding perturbations to inputs, preserving graph structure while evaluating f.

Experimental results

Research questions

• RQ1 Can subgraph-level explanations provide more intuitive and faithful insights than node/edge explanations for GNNs?
• RQ2 How can subgraph importance be quantified to account for interactions between graph components?
• RQ3 What is the trade-off between explanation fidelity and computational efficiency when using SubgraphX with Shapley-based scoring?
• RQ4 How well does SubgraphX perform across graph classification and node/link prediction tasks?
• RQ5 What are practical approximations to make Shapley-value computation scalable on real graphs?

Key findings

• SubgraphX yields more human-intelligible explanations by identifying connected subgraphs rather than isolated edges or nodes.
• Shapley-value scoring captures interactions among subgraphs and yields faithful explanations.
• With approximations (L-hop pruning and Monte Carlo coalition sampling), SubgraphX achieves competitive fidelity with substantially lower cost than exact Shapley computation.
• On graphs like BA-2Motifs and MUTAG, SubgraphX localizes the key substructures corresponding to ground-truth motifs or chemistry-relevant groups.
• Across datasets and models (GCN/GAT/GIN), SubgraphX outperforms baselines in fidelity under matched sparsity levels, with reasonable runtime.
• SubgraphX explanations remain applicable to graph classification, node classification, and link prediction tasks.