Heterogeneous Dynamic Dependency Graph Learning for Anomaly Detection in Large-Scale Backend Systems

Abstract

This paper addresses the anomaly detection needs of microservice systems characterized by continuously evolving dependencies, diverse call semantics, and heterogeneous, multi-source observation data. It constructs a dynamic dependency anomaly detection framework based on heterogeneous graph neural networks. The method represents the system runtime as a time-organized heterogeneous dynamic graph, explicitly characterizing service entities, interaction relationships, and their type semantics within a unified structure. Observational features such as logs, metrics, and links are mapped to a shared representation space to achieve cross-source alignment. To enhance the characterization of dependency reorganization and propagation effects, the framework employs a message passing and attention aggregation mechanism under relation type constraints, selectively converging information along semantically consistent dependency channels, thereby strengthening the contribution of key interactions to anomaly detection. For time-series non-stationarity and short-term disturbances in dynamic environments, the method introduces learnable temporal representations and adjusts the intensity of temporal information injection through gating fusion, reducing noise amplification and enhancing the stability of state representation. Furthermore, the framework constructs edge representations at the relation layer and outputs anomaly scores to consider the joint discrimination of node states and dependency semantics. Comparative experimental results show that the proposed method outperforms other methods on multiple evaluation metrics, verifying the effectiveness and robustness of heterogeneous dependency modeling and temporal information fusion in dynamic dependency anomaly detection tasks.