Graph Learning Framework for Precise Anomaly Localization in Distributed Microservice Environments

Abstract

 This paper proposes an automatic anomaly call chain localization method based on graph neural networks to address the challenge of anomaly detection in modern microservice systems. The method constructs a directed graph using the invocation relationships between microservices. The operational features of service nodes are encoded as node attributes. A graph neural network is then used to model the structure of the call chain in depth, enabling precise identification of potential anomaly nodes. Unlike traditional methods that rely on static rules or single-point metric analysis, the proposed model incorporates structural awareness and multi-hop information aggregation. This allows it to effectively capture the structural patterns of anomaly propagation within service call paths. To evaluate the model's performance, a series of experiments were conducted. These assess the model under different conditions, including feature combination strategies, system load intensity, anomaly type diversity, and propagation path depth. The experimental results show that the proposed method achieves significantly better performance than representative public methods in recent literature, across key metrics such as F1 Score, Precision, and Recall. The model demonstrates higher accuracy and stability. Furthermore, the method maintains strong identification capability in handling complex anomaly propagation patterns, high-concurrency loads, and multi-level service chains. This highlights its practical applicability in real engineering environments. The proposed method provides an effective modeling framework for intelligent anomaly localization in large-scale microservice systems. It also advances the application of graph learning techniques in system operations and maintenance.