Interpreting Graph Inference with Skyline Explanations

TL;DR

This work tackles the challenge of interpreting GNN predictions by proposing skyline explanations, a Pareto-optimal framework that jointly optimizes multiple explainability criteria with respect to a measure space $\Phi$. It introduces Skyline Explanatory Query ($\mathsf{SXQ}$) to compute a non-dominated set of explanatory subgraphs, proves hardness results for exact evaluation, and delivers practical, scalable algorithms: onion-peeling based ASX-OP, edge-growing ASX-I, and diversification-driven DSX, along with a parallel querying framework ParaSX. The methods provide provable guarantees (e.g., $(\frac{1}{4},\epsilon)$- and $(\frac{1}{2}-\epsilon)$-approximation) and demonstrate superior multi-criteria explanation quality and scalability on large real-world graphs, outperforming existing GNN explainers across multiple datasets and measures. The work highlights the practical impact of providing diversified, multi-faceted interpretations for GNN decisions, enabling more trustworthy and comprehensive decision support.

Abstract

Inference queries have been routinely issued to graph machine learning models such as graph neural networks (GNNs) for various network analytical tasks. Nevertheless, GNN outputs are often hard to interpret comprehensively. Existing methods typically conform to individual pre-defined explainability measures (such as fidelity), which often leads to biased, ``one-side'' interpretations. This paper introduces skyline explanation, a new paradigm that interprets GNN outputs by simultaneously optimizing multiple explainability measures of users' interests. (1) We propose skyline explanations as a Pareto set of explanatory subgraphs that dominate others over multiple explanatory measures. We formulate skyline explanation as a multi-criteria optimization problem, and establish its hardness results. (2) We design efficient algorithms with an onion-peeling approach, which strategically prioritizes nodes and removes unpromising edges to incrementally assemble skyline explanations. (3) We also develop an algorithm to diversify the skyline explanations to enrich the comprehensive interpretation. (4) We introduce efficient parallel algorithms with load-balancing strategies to scale skyline explanation for large-scale GNN-based inference. Using real-world and synthetic graphs, we experimentally verify our algorithms' effectiveness and scalability.

Figures (9)

• Figure 1: A Bitcoin transaction network with a target IP address (test node $v_t$) that has a label " illicit" to be explained ellipticpp.
• Figure 2: Illustration of Onion Peeling ($L$=$2$, $k$=$2$). $\zeta'$ = $\{s_1, s_2, s_3, s_5\}$. $(\zeta',\epsilon)$-explanation ${\mathcal{G}}_\epsilon$=$\{s_2, s_5\}$ with ${\sf DS}$ = $4$.
• Figure 3: Effectiveness of our skyline explanation vs. baseline explainers ($$GExp, $$PGExp, $$CF2, $$MOExp, $$SAME).
• Figure 4: Efficiency and scalability of our skyline explanation vs. SOTA explainers ($$GExp, $$PGExp, $$CF2, $$MOExp, $$SAME).
• Figure 5: Qualitative analysis of our skyline explanation and baseline explainers ($$GExp, $$PGExp, $$CF2, $$MOExp, $$SAME).