Research on a Graph Contrastive Learning-Based Anomaly Detection Method for Backend Microservice Call Graph Structure Change Awareness

Authors
  • Yuruo Gong

    Rice University, Houston, USA

    Author

  • Arjun Mehta

    University of Southern California, Los Angeles, USA

    Author

  • Zixuan Zhou

    University of Southern California, Los Angeles, USA

    Author

Abstract

Addressing the challenges of complex call relationships, dynamic changes in runtime states, and hidden anomaly propagation paths in backend microservice systems, this paper proposes a graph-based contrastive learning anomaly detection method that is aware of call graph structure changes. This method constructs a time-varying call graph based on the microservice call chain, incorporating service node states, dependencies, and their cross-time window structural changes into a unified modeling framework. It introduces a graph difference representation mechanism to characterize the evolutionary information between call graphs at adjacent time points and combines graph encoding, temporal aggregation, view generation, and change-guided representation learning to enhance the model's ability to distinguish between normal evolutionary behavior and anomalous perturbation patterns. Furthermore, it utilizes change-aware attention and adaptive message passing mechanisms to enhance the representation weights of key nodes and important dependencies, enabling the model to more effectively capture potential anomaly features such as local structural imbalances, abnormal link reconstruction, and service interaction offsets. Compared to methods relying solely on static topology or single-indicator modeling, the proposed method more fully mines the dynamic correlation information and structural semantics within the microservice system, thereby improving the accuracy and stability of anomaly identification. Comparative experimental results show that the proposed method exhibits good overall performance in anomaly detection tasks, verifying the effectiveness of combining structural change perception with graph contrastive learning, and demonstrating that backend microservice anomaly modeling from the perspective of time-varying call graphs has strong feasibility and application value.

Cover Image
Downloads
Published
2024-06-01
Section
Articles