Structure-Aware Root Cause Localization for Microservice Backends via Joint Graph Representation Learning and Causal Inference

Abstract

Service oriented architectures improve system flexibility and scalability through decoupling and independent deployment. At the same time, they introduce complex dependencies, hidden anomaly propagation paths, and highly heterogeneous observations. These factors make backend fault root cause localization particularly challenging. In microservice systems, anomalies often cascade along dependency structures, and surface symptoms are separated from true triggers in time and space. To address these characteristics, this paper proposes a backend fault root cause localization method based on microservice dependency graphs. Graph representation learning and causal inference are unified within a single modeling framework. Service invocations and resource dependencies are used to construct a system dependency graph. Logs, metrics, and traces are mapped into node state representations. System-level runtime states are learned under structural constraints. By modeling service interactions on the graph, the method enables system-level understanding of anomaly organization rather than decisions based on local signals alone. Causal modeling is further introduced to capture directional influences among service nodes. Anomaly analysis is thus elevated from correlation-based description to a mechanism-level explanation. This design allows a clear distinction between triggering sources and propagation effects. Comparative analysis under the same evaluation setting shows that the proposed method achieves more stable performance in overall discrimination, result consistency, and system-level anomaly characterization. This work provides a unified modeling approach that is structure-aware and causally interpretable. It improves the reliability of anomaly understanding and diagnosis in complex microservice backend systems.