When Logs Are No Longer Just Sequences: A Multi-view Bidirectional Graph Convolutional Network for Anomaly Detection

Abstract

Industrial information service systems are becoming increasingly complex, and timely anomaly detection and diagnosis are crucial for ensuring the quality of software services. System logs constitute a direct and important source of information for anomaly detection and diagnosis. However, existing methods struggle to simultaneously model local and global dependencies, and they provide insufficient representations of directional semantics and causal chains in log execution sequences. To overcome these limitations, we propose a Multi-view Bidirectional Graph Convolutional Network (BiView-GCN), which employs a collaborative representation mechanism combining partial bidirectional graph convolution and global graph diffusion. This mechanism enhances anomaly discrimination while more comprehensively capturing directional information and complex dependencies in log execution sequences. Crucially, BiView-GCN decomposes anomaly scores at the node and edge levels, enabling fine-grained attribution. Experimental results demonstrate that our method not only achieves superior performance but also provides node- and edge-level attribution results, offering clues for subsequent anomaly diagnosis.