高阶结构增强的跨视图无负样本对比的图异常检测算法

Cross-View Negative-Free Contrastive Learning for Graph Anomaly Detection with High-Order Structure Augmentation

图异常检测在网络安全、金融评估和医疗保健等多个领域都有广泛的实际应用。近年来,基于对比学习和基于生成重构的图异常检测算法框架取得了显著的性能提升。然而,大多数基于图神经网络的范式忽略了一个内在的缺点,即可能会无意识地将异常节点与其邻域正常节点聚合在一起。此外,现有的检测算法缺乏对高阶结构信息的关注,导致正常节点与异常节点之间的判别性下降。为了改善以上缺点,提出了一种高阶结构增强的跨视图无负样本对比的图异常检测算法(CNCL-GAD)。与现有的单视图对比范式不同,提出了以高阶结构信息作为增强视图,通过多视图对比学习方法为图异常检测任务(GAD)引入更多、更丰富的判别信息。为了缓解图异常检测任务中正常样本与异常样本类别不平衡导致的对比负样本对大多数是同一类别的现象,提出了跨视图无负样本对比策略,即只将两个视图之间的正样本子图对拉近。将视图内节点-子图对比模块、属性重构模块和跨视图子图-子图对比模块联合训练,以获得更好的检测性能。在现有的公开数据集上进行了大量实验,与其他竞争算法相比,所提出的算法实现了有竞争力甚至更优越的性能。

Graph anomaly detection has practical applications in various fields,such as cyber security,financial eval-uation and medical care.Recently,contrastive-based and generative-based detection frameworks have achieved re-markable performance improvements.However,most of the existing paradigms overlook the drawback that the GCN-based framework may unconsciously aggregate abnormal nodes with their neighborhood normal partners.Moreover,these detection algorithms lack attention to high-order structural information.These lead to a reduction in the distinction between normal nodes and their opponents.To bridge the gaps above,this paper proposes a cross-view negative-free contrastive learning utilizing high-order structure for graph anomaly detection(CNCL-GAD)in this paper.Especially,different from the existing single-view contrastive paradigm,this paper develops the high-order structure as the augmented view to introduce more global abnormality discrimination with multi-view contrastive learning for graph anomaly detection(GAD).Then,to mitigate the false-negative phenomenon of imbalanced data in GAD tasks where the majority of selected contrastive negative samples are normal subgraphs,this paper proposes the cross-view negative-free contrastive strategy to only pull the positive subgraphs’pairs between two views as close as possible.Furthermore,this paper integrates intra-view node-subgraph contrastive modules,attribute recon-struction modules,and cross-view subgraph-subgraph contrastive modules to simultaneously obtain more distinc-tions on structure and attribute.The extensive experiments conducted on benchmark datasets show that the proposed method achieves competitive or even superior performance compared with existing competitors.