Fine-Grained Backend Anomaly Classification through Structural Dependency Learning and Version-Aware Feature Alignment

Abstract

This study addresses the challenge of anomaly identification in backend systems operating under highly dynamic access environments and frequent version evolution. It proposes a unified anomaly classification framework that integrates access pattern modeling, structural dependency representation, and version modulation. The method first constructs high-dimensional temporal features to capture local variations and global trends in access behaviors. It then applies graph structure learning to describe coupling relationships among indicators and link-level association patterns, enabling the model to understand potential dependencies in backend operations. A controllable version embedding is introduced to model behavioral differences across iterations and to reduce distribution drift caused by system evolution. To ensure stable and separable final representations, the method incorporates feature alignment constraints and structural boundary constraints, which enhance the separability of anomaly categories in the latent space. Based on this joint representation, a classifier performs fine-grained discrimination across multiple anomaly types. The experiments include model comparisons and sensitivity studies on hyperparameters, environments, and data, providing a comprehensive evaluation of the framework's adaptability and robustness in complex access scenarios. The results show that the proposed method achieves significant advantages in accuracy, stability, and cross-version consistency and effectively captures fine-grained anomalies in backend operations. This anomaly classification framework offers a structured solution for automated anomaly identification in highly dynamic backend environments and establishes a technical foundation for improving system stability.