Adaptive Fusion of Traces and Logs via Graph Attention for Microservice Anomaly Localization

Abstract

 Microservice architectures have become increasingly prevalent in modern cloud-native  applications due to their flexibility and scalability benefits. However, the distributed nature of  microservices introduces significant challenges in anomaly detection and root cause localization,  as failures can propagate across multiple services and manifest differently in various monitoring  data sources. Traditional approaches that rely on single-modal data such as metrics, logs, or  traces often fail to capture the complete picture of system behavior, leading to high false positive  rates and missed anomalies. This paper proposes an adaptive fusion framework that integrates  distributed traces and application logs through Graph Attention Networks (GAT) to achieve  accurate microservice anomaly localization. The proposed approach constructs a unified graph  representation that captures both the structural dependencies between services from traces and  the semantic information from logs. By leveraging the attention mechanism, our method adaptively  assigns importance weights to different data modalities and service interactions, enabling the  model to focus on critical anomaly patterns while suppressing noise. Experimental evaluation on  benchmark microservice systems demonstrates that the proposed framework achieves superior  performance in both anomaly detection and root cause localization compared to existing single modal and naive multi-modal approaches, with detection accuracy exceeding 92% and  localization precision reaching 87%.