摘要
近年来,小样本图异常检测在各个领域中引起了广泛的研究兴趣,其旨在在少量有标记训练节点(支持集)的引导下去检测出大量无标记测试节点(查询集)中的异常行为。然而,现有的小样本图异常检测算法通常假设其可以从具有大量有标记节点的训练任务(元训练任务)中学习,从而有效地推广到具有少量标记节点的测试任务(元测试任务),这一假设并不符合真实世界的应用条件。在实际应用中,用于小样本图异常检测训练的元训练任务通常只包含极其有限的有标记节点,其标签占比通常不超过0.1%,甚至更低。由于元训练和元测试任务之间存在的巨大任务差异,现有的小样本图异常检测算法很容易出现模型的过拟合问题。除此之外,现有的小样本图异常检测算法仅利用节点间的一阶邻域(局部结构信息)来学习节点的低维特征嵌入,反而忽略了节点间的长距离依赖关系(全局结构信息),进而导致学习到的低维特征嵌入的不准确性和失真问题。针对上述挑战,本文提出了极其弱监督场景下的小样本图异常检测算法——EWSFSGAD。具体来说,该方法首先提出了一个简单且有效的图神经网络框架——GLN(Global and Local Network),其能够同时有效地利用节点间的全局和局部结构信息,并进一步引入注意力机制实现节点间的信息交互,从而更加有效地学习节点鲁棒的低维特征嵌入;该方法还引入了图对比学习中的自监督重建损失,使得节点原始视图与其增强视图之间低维特征嵌入的互信息尽可能一致,为EWS-FSGAD模型的优化提供更多有效的自监督信息,进而提升模型的泛化性;为了提升模型在真实场景中小样本图异常检测任务的快速适应性,该方法引入跨网络元学习训练机制,从多个辅助网络学习可迁移元知识,为模型提供良好的参数初始化,从而能够通过在仅有很少甚至一个标记节点的目标网络上进行微调并有效泛化。在三个真实世界的数据集(Flickr、PubMed、Yelp)上的大量实验结果表明,本文所提方法的性能明显优于现有的图异常检测算法。特别是在PubMed数据集上,AUC-PR提升了28.8%~35.4%。这些实验结果强有力地证明了在极其有限标记的元训练任务引导下,本文所提方法能够更好地学习到异常节点本质特征,从而提升小样本图异常检测任务的有效性。
In recent years,few-shot graph anomaly detection(FS-GAD)has received extensive research interest across various applications,which aims to distinguish anomalous patterns of abundant unlabeled test nodes(query set)under the guidance of a few labeled training nodes(support set).Nevertheless,the existing FS-GAD methods often assume that they can learn metaknowledge from training tasks(meta-training tasks)with abundant labeled nodes,and then effectively generalize to testing tasks(meta-test tasks)with a few labeled nodes.This assumption does not fit with real-world applications.In real-world applications,the meta-training tasks for FSGAD training usually contain only extremely limited labeled nodes,whose proportion of labels usually does not exceed 0.1%or even less.Owing to the large task difference between metatraining and meta-testing tasks,the existing FS-GAD methods are more prone to overfitting problems.In addition,the existing FS-GAD methods only utilize the first-order neighborhood(local structure information)between nodes to learn their low-dimensional node feature embedding,while ignoring the long-range dependencies(global structure information)between nodes,leading to the inaccuracy and distortion of the learned low-dimensional node feature embeddings.In this paper,an effective few-shot graph anomaly detection framework with extremely weak supervision is proposed,termed EWS-FSGAD,to solve the above-mentioned issues.Specifically,a simple and effective graph neural network module~Global and Local Network(GLN)is first proposed to more effectively learn robust low-dimensional node feature embeddings,which simultaneously utilizes the global and local structural information between nodes and also introduces the attention mechanism to realize the information interaction between nodes.And then,we introduce self-supervised reconstruction loss in graph contrast learning to maximize the mutual information between lowdimensional node feature embeddings from the original view and the augmented view,which can provide more effective self-supervised information for model optimization and also further improve the generalization of the proposed EWS-FSGAD.To improve the effectiveness of the proposed EWS-FSGAD method in real-world applications,we introduce the cross-network meta-learning training mechanism to learn transferable meta-knowledge from multiple auxiliary networks and provide good parameters initialization for the proposed EWS-FSGAD model,so that it can quickly adapt to the target network by performing fine-tuning on a few or even one labeled node.Finally,extensive experiments on three real-world benchmarks(Flickr,PubMed,and Yelp)show that the proposed EWS-FSGAD achieves state-of-the-art performance in comparison to the existing graph anomaly detection models.For example,AUC-PR improves by 28.8%~35.4%on the PubMed dataset.These results strongly demonstrate that the proposed EWS-FSGAD can better learn the essential characteristics of abnormal nodes under the guidance of an extremely limited meta-training support set,and further improve the effectiveness of FS-GAD tasks.
作者
郑文捷
傅司超
陈嘉真
彭勤牧
涂益群
邹斌
荆晓远
尤新革
ZHENG Wen-Jie;FU Si-Chao;CHEN Jia-Zhen;PENG Qin-Mu;TU Yi-Qun;ZOU Bin;JING Xiao-Yuan;YOU Xin-Ge(School of Electronic Information and Communications,Huazhong University of Science and Technology,Wuhan 430074;Department of Statistics and Actuarial Science,University of Waterloo,Waterloo ON N2L 3G1,Canada;Huaneng Wuhan Power Generation Co.Ltd.,Wuhan 430050;Faculty of Mathematics and Statistics,Hubei University,Wuhan 430062;School of Computer Science,Wuhan University,Wuhan 430072)
出处
《计算机学报》
北大核心
2025年第4期927-948,共22页
Chinese Journal of Computers
基金
国家重点研发计划(2022YFF0712300)
中央高校基本科研业务费(YCJJ20241203)资助。
关键词
图异常检测
小样本学习
极其弱监督
图神经网络
图对比学习
长距离依赖关系
graph anomaly detection
few-shot learning
extremely weak supervision
graph neural networks
graph contrastive learning
long-range dependency relationships
