Page 264 - 《软件学报》2026年第7期

岳立楠 等: 面向图分类任务的互补感知证据提取方法 2949

解耦机制和回声学习策略的有效性.

5 总 结

本文提出了一种面向图分类任务的互补感知证据提取方法 CaR, 旨在通过引入反事实数据增强策略, 有效提
升图神经网络模型的可解释性与泛化能力. 该方法基于一个核心假设: 在给定证据的条件下, 互补部分与标签应保
持相互独立. 该假设的引入有助于构造更具判别性的反事实样本, 从而缓解模型对伪相关特征的依赖问题. 为实现
上述目标, 本文首先设计了基于对比学习的证据-互补解耦机制, 从表示空间层面促使互补表示远离证据表示, 增
强了互补部分与标签之间的独立性, 有效提升了模型对因果结构的辨识能力. 进一步地, 针对现有方法普遍存在仅
依赖最终消息传递层获取互补表示, 导致互补信息可能包含证据信息的问题, 本文提出回声学习 (echo-learning)
策略. 该策略通过在多个早期图神经网络层中采样互补表示, 显著提升了互补信息的多样性, 从而增强了反事实样
本的表达能力, 提升了模型在复杂图结构下的适应能力与泛化能力. 在多个真实图分类数据集与构造的合成数据
集上进行的系统实验证明, 本文所提 CaR 方法在任务预测性能与证据提取准确性等方面均显著优于现有代表性
方法, 说明了提出方法的有效性.
尽管本文提出的回声学习与互补感知反事实增强策略在缓解捷径学习和提升解释忠实性方面取得了良好效
果, 但仍存在一定局限性. 首先, 回声学习中历史层的选择及融合权重主要依赖经验设定, 不同数据集和网络深度
下的最优层数可能存在差异, 尚缺乏自适应的层选择机制, 影响方法的通用性与可复现性. 其次, 虽然通过打乱或
重组互补表示能够有效破坏证据与环境之间的捷径共现关系, 从而削弱模型对非因果相关性的依赖, 但所构造的
反事实样本在部分情况下可能偏离真实数据分布, 可能对训练过程的稳定性以及模型在真实应用场景中的泛化性
能产生一定影响, 有待在后续工作中结合分布约束或生成式建模进一步加以缓解.

References
[1] Hu WH, Fey M, Zitnik M, Dong YX, Ren HY, Liu BW, Catasta M, Leskovec J. Open graph benchmark: Datasets for machine learning
on graphs. In: Proc. of the 34th Int’l Conf. on Neural Information Processing Systems. Vancouver: Curran Associates Inc., 2020. 1855.
[2] Guo ZC, Zhang CX, Yu WH, Herr J, Wiest O, Jiang M, Chawla NV. Few-shot graph learning for molecular property prediction. In: Proc.
of the 2021 Web Conf. Ljubljana: ACM, 2021. 2559–2567. [doi: 10.1145/3442381.3450112]
[3] Yu B, Yin HT, Zhu ZX. Spatio-temporal graph convolutional networks: A deep learning framework for traffic forecasting. In: Proc. of the
27th Int’l Joint Conf. on Artificial Intelligence. Stockholm: AAAI Press, 2018. 3634–3640.
[4] Wu SW, Sun F, Zhang WT, Xie X, Cui B. Graph neural networks in recommender systems: A survey. ACM Computing Surveys, 2023,
55(5): 97. [doi: 10.1145/3535101]
[5] Ying R, Bourgeois D, You JX, Zitnik M, Leskovec J. GNNExplainer: Generating explanations for graph neural networks. In: Proc. of the
33rd Int’l Conf. on Neural Information Processing Systems. Vancouver: Curran Associates Inc., 2019. 829.
[6] Fang JF, Liu W, Gao Y, Liu ZM, Zhang A, Wang X, He XN. Evaluating post-hoc explanations for graph neural networks via robustness
analysis. In: Proc. of the 37th Int’l Conf. on Neural Information Processing Systems. New Orleans: Curran Associates Inc., 2024. 3168.
[7] Lee S, Wang XT, Han S, Yi XY, Xie X, Cha M. Self-explaining deep models with logic rule reasoning. In: Proc. of the 36th Int’l Conf.
on Neural Information Processing Systems. New Orleans: Curran Associates Inc., 2022. 232.
[8] Gui SR, Liu M, Li XN, Luo YZ, Ji SW. Joint learning of label and environment causal independence for graph out-of-distribution
generalization. In: Proc. of the 37th Int’l Conf. on Neural Information Processing Systems. New Orleans: Curran Associates Inc., 2024.
174.
[9] Sui YD, Tang CZ, Chu ZX, Fang JF, Gao Y, Cui Q, Li LF, Zhou J, Wang X. Invariant graph learning for causal effect estimation. In:
Proc. of the 2024 ACM Web Conf. Singapore: ACM, 2024. 2552–2562. [doi: 10.1145/3589334.3645549]
[10] Wu YX, Wang X, Zhang A, He XN, Chua TS. Discovering invariant rationales for graph neural networks. In: Proc. of the 10th Int’l Conf.
on Learning Representations. OpenReview.net, 2022.
[11] Sui YD, Wang X, Wu JC, Lin M, He XN, Chua TS. Causal attention for interpretable and generalizable graph classification. In: Proc. of
the 28th ACM SIGKDD Conf. on Knowledge Discovery and Data Mining. Washington: ACM, 2022. 1696–1705. [doi: 10.1145/3534678.
3539366]
[12] Yue LN, Liu Q, Liu Y, Gao WB, Yao FZ, Li WF. Cooperative classification and rationalization for graph generalization. In: Proc. of the