Learning Unified Multi-Granularity Representations for Backend Anomaly Detection and Causal Localization

Abstract

As backend systems grow in scale and architecture complexity, performance anomalies often exhibit characteristics such as multi-source signal coupling, cross-component propagation, and varied morphologies, making it difficult for solutions relying solely on single observations or single-granularity modeling to balance accuracy and interpretability. This paper proposes a unified framework based on self-supervised representation learning for multi-granularity backend performance anomaly detection and root cause inference. This framework aligns and normalizes heterogeneous telemetry data such as metrics, logs, and tracking within the same time window, and learns stable latent representations through a shared encoder. In terms of method design, the framework simultaneously characterizes temporal context and dependency structure information: on the one hand, it extracts window-level states using temporal aggregation; on the other hand, it constructs a graph structure based on call relationships and performs relational representation fusion. Subsequently, a self-supervised objective with consistency constraints is adopted to improve the robustness of the representation to noise and scene fluctuations. In the inference stage, this paper constructs an anomaly scoring mechanism in the latent space and combines neighborhood consistency for score propagation and aggregation, thereby outputting root cause ranking results that can be used for localization. Through systematic comparison on public benchmarks, the proposed method demonstrates stronger overall advantages in both detection quality and root cause hit capability, verifying the effectiveness and practical value of multi-source telemetry alignment, multi-granularity representation learning, and structural consistency inference in backend anomaly diagnosis tasks.