윤예진, 임재성, and 김지혜 publish a paper at CIKM 2026

Context-Aware Subgraph Explanations for Multi-Task GNNs was published at CIKM 2026. This study proposes CAMIE, a new method for explaining the predictions of multi-task molecular graph neural networks (GNNs). Existing GNN explanation methods primarily search for important subgraphs for individual tasks and do not fully account for the possibility that, even for the same molecule, the important substructures (motifs) may vary with the biological property being predicted (the assay). To address this, the proposed method estimates task-specific motif importance using both structural information about the molecule and semantic information about the assay.

The method uses knowledge distillation to learn how predictions from a pretrained GNN change when a particular motif is removed. This enables a single shared model to learn explanations for multiple tasks without separate ground-truth explanations. Experiments on ToxCast data showed greater explanation faithfulness than existing methods, particularly for tasks with limited training data. The results demonstrate that multi-task GNN explanations can share information across tasks while reflecting the characteristics of each task.