Modeling Evolving Service Dependencies: Dynamic Graph Learning for Microservice Anomaly Detection

Abstract

Addressing the challenges of anomaly detection in cloud-native microservice systems due to frequent changes in backend dependencies and strong time-varying coupling of runtime states, this paper investigates the problem of dynamic graph temporal anomaly detection and proposes a joint structural and temporal modeling framework. The method organizes the service dependency graph using time-slice sequences, unifying the representation of service nodes, call relationships, and their multi-source attributes. It encodes dependency interaction strength and local structural consistency through relation-aware neighborhood aggregation, and uses a temporal aggregation module to characterize the cross-time state evolution patterns and long-term dependencies. Based on conditional prediction of normal evolution trajectories, an anomaly metric centered on prediction residuals is constructed, transforming the impact of structural perturbations and behavioral drift on the representation space into quantifiable anomaly scores, thereby supporting node-level and subgraph-level anomaly detection and event characterization. Experiments are conducted on open-source microservice monitoring data, and multiple evaluation metrics verify the effectiveness of the proposed method in terms of anomaly discrimination capability and overall recognition quality.