| 摘要: |
| 图神经网络(graph neural network, GNN)在图分类任务中表现出色, 但其“黑箱”性质引发了对预测过程可解释性的广泛关注. 作为一种自解释机制, 证据提取方法近年来受到广泛研究, 其目标是在生成预测结果的同时, 从原始图中提取紧凑、具备判别性的子图结构(即证据子图)以作为解释. 然而, 现有方法往往依赖数据中的“捷径”特征进行预测, 导致提取的解释缺乏忠实性, 进而影响模型的可解释性与鲁棒性. 为缓解上述问题, 提出一种互补感知证据提取(complement-aware rationale extraction, CaR)方法, 将未被选为证据的子图区域视为互补信息, 并从以下3个方面提升反事实建模与解释能力: 首先, 引入对比学习机制以解耦证据表示与互补表示, 增强二者的语义独立性; 其次, 提出回声学习(echo-learning)策略, 充分利用GNN各层消息传递过程中的中间表示, 捕捉不同深度层次下互补结构的差异性; 最后, 将当前与历史层的互补表示与证据表示结合, 进而构造反事实样本, 提升训练数据的多样性. 在多个真实数据集及一个合成数据集上的实验证明, CaR能够生成更具忠实性的证据, 验证了其有效性. |
| 关键词: 证据提取 图神经网络 捷径学习 可解释人工智能 数据挖掘 |
| DOI：10.13328/j.cnki.jos.007655 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(62506072, 62225602) |
|
| Complement-aware Rationale Extraction Method for Graph Classification Tasks |
|
YUE Li-Nan1,2, ZHANG Min-Ling1,2
|
|
1.School of Computer Science and Engineering, Southeast University, Nanjing 211189, China;2.Key Laboratory of Computer Network and Information Integration (Southeast University), Ministry of Education, Nanjing 211189, China
|
| Abstract: |
| Graph neural networks (GNNs) have achieved remarkable performance on graph classification tasks, but their black-box nature has raised widespread concerns about the explainability of their prediction process. As a self-explaining mechanism, rationale extraction has received increasing attention in recent years. Its goal is to extract concise subgraph structures from the original graph (i.e., rationale subgraphs) as explanations while generating prediction results. However, existing methods often rely on spurious shortcut features in the data, resulting in explanations that lack faithfulness, which in turn compromises both the interpretability and robustness of the model. To address these issues, this study proposes a complement-aware rationale extraction (CaR) method, which treats the subgraph regions not selected as rationales as complement information. The method enhances counterfactual modeling and interpretability from the following three perspectives. First, a contrastive learning mechanism is introduced to disentangle rationale representations from complement representations, enhancing their semantic independence. Second, an echo-learning strategy is proposed to fully leverage the intermediate representations generated during the message-passing process of GNNs, capturing the structural differences in complement parts across different network depths. Finally, the method combines complement and rationale representations from both current and historical layers to construct counterfactual samples, increasing the diversity of the training data. Extensive experiments on multiple real-world benchmark datasets and a synthetic dataset demonstrate the effectiveness of CaR in producing faithful rationales. |
| Key words: rationale extraction graph neural network (GNN) shortcut learning explainable AI data mining |
