Context-Aware Subgraph Explanations for Multi-Task GNNs
TL;DR
The proposed CAMIE, a Context-Aware Motif Importance Explainer designed for multi-task molecular classification, leverages a Knowledge Distillation-inspired pseudo target to effectively distill assay-specific motif-removal responses, transferring explanatory patterns across tasks without requiring ground-truth labels.
Abstract
Explanations for Graph Neural Networks (GNNs) that perform graph-level prediction generally describe how substructures in an input graph contribute to a specific output. However, in multi-task settings, the input-output relationship and important substructures are heterogeneous across tasks, limiting the effectiveness of existing explanation methods. Existing approaches typically relied on task-specific independent explainers, underutilizing the shared representations across related tasks. As a result, it remains challenging to transfer explanatory patterns to data-scarce tasks. In this paper, we propose CAMIE, a Context-Aware Motif Importance Explainer designed for multi-task molecular classification. CAMIE leverages a Knowledge Distillation-inspired pseudo target to effectively distill assay-specific motif-removal responses, transferring explanatory patterns across tasks without requiring ground-truth labels. CAMIE utilizes assay context information as an input to the explainer, jointly capturing shared explanatory structures and task-dependent mechanisms. Our experiments on the ToxCast dataset show that our proposed method flexibly adapts explanations according to assay context and provides more faithful explanations across tasks, including tasks with limited data.
