| 摘要: |
| 大多数图异常检测方法依赖图神经网络(GNN)在相对高质量的图数据上进行学习. 然而, 在现实应用中, 这种理想场景极为罕见, 大多数数据存在标签缺失、动态变化和结构不完整等问题, 这些问题统称为动态不完整图. 针对GNN在极端条件下性能下降的挑战, 提出一种可解释的动态不完整图异常检测方法EXDIG (explainable dynamic incomplete graph anomaly detection), 其核心是一种结合强-弱互信息优化的图掩码自编码器框架. 该框架通过对图结构(节点/边)和节点特征进行掩码, 模拟现实中的动态不完整场景. 此外, 通过强-弱互信息损失, EXDIG捕捉结构与特征之间的关系, 同时保持结构完整性, 降低过拟合风险, 并提升泛化能力. 此外, 该方法通过在节点、边及特征上引入掩码扰动, 提高动态不完整图异常检测的可解释性, 使其能够识别关键组成部分, 并为异常检测结果提供透明且可信的解释. 在9个真实世界图数据集上进行评估, 实验结果表明, EXDIG在不同程度的动态不完整场景下, 在多种下游任务和表示学习评估(包括有监督和无监督设定)中均优于现有最先进方法. 其中, 在异常检测数据集Amazon上, EXDIG的NMI和ARI指标分别提升了超过 13% 和 15%; 在动态不完整比率从25%到99%的设置下, 其F1分数波动被控制在5%以内. 此外, EXDIG还实现了在动态不完整图中对各节点的可解释性分析. |
| 关键词: 可解释的图学习 掩码自编码器 互信息 图异常检测 |
| DOI：10.13328/j.cnki.jos.007523 |
| 分类号:TP18 |
| 基金项目:国家自然科学基金(62302287) |
|
| Explainable Dynamic Incomplete Graph Anomaly Detection Based on Masked Learning with Strong-weak Mutual Information |
|
LUO Xiang-Feng, GU Jun-Quan, YU Hang
|
|
School of Computer Engineering and Science, Shanghai University, Shanghai 200444, China
|
| Abstract: |
| Most graph anomaly detection methods leverage graph neural network (GNN) to learn from relatively high-quality graph data. Unfortunately, such ideal scenarios are rare in real-world applications, where most data suffer from issues such as missing labels, dynamic changes, and structural incompleteness, collectively referred to as dynamic incomplete graph (DIG). To address the challenge of performance degradation of GNN under extreme conditions, this study proposes an explainable dynamic incomplete graph anomaly detection (EXDIG) method. The core is a graph masked autoencoder framework optimized with strong-weak mutual information. This framework simulates real-world DIG scenarios by masking graph structures (nodes/edges) and node features. In addition, through the strong-weak mutual information (SWMI) loss, it captures the relationship between structure and features while maintaining structural integrity, reducing overfitting, and improving generalization. Furthermore, EXDIG enhances the interpretability of anomaly detection in DIGs by incorporating masked perturbations on nodes, edges, and features, enabling the identification of key components and providing transparent, trustworthy explanations for anomaly detection results. This study evaluates EXDIG on nine real-world graph datasets, and the results demonstrate its superiority over state-of-the-art methods across different levels of DIG scenarios and across various downstream tasks and representation learning evaluations, both supervised and unsupervised. Specifically, on the Amazon anomaly detection dataset, EXDIG achieves improvements over 13% and 15% in NMI and ARI, respectively. It maintains F1-score fluctuations within 5% across dynamic incompleteness ratios from 25% to 99%. Notably, EXDIG enables node-level interpretability in dynamic incomplete graphs. |
| Key words: explainable graph learning masked autoencoder mutual information graph anomaly detection |
