Modeling Complex Service Dependencies for Multidimensional Anomaly Detection via Attention Mechanisms

Abstract

This study proposes an attention-driven anomaly detection model for complex service dependencies to address the challenges of dynamic coupling among multidimensional metrics, intricate dependency structures, and diverse anomaly patterns in cloud service systems. The model achieves unified representation of temporal dynamics and structural semantics through multi-scale temporal feature encoding and an adaptive dependency modeling mechanism. In the feature extraction stage, a multi-head attention mechanism captures both local fluctuations and long-term trends across multi-granularity time windows, enhancing semantic interaction and contextual awareness among features. In the structural modeling stage, dynamic graph construction and dependency sparsification strategies are employed to characterize latent semantic associations and propagation relationships among services, enabling effective identification of cross-node anomaly patterns. Furthermore, a temporal consistency regularization term is introduced to maintain cross-time smoothness and dependency continuity of latent states, ensuring robustness and generalization under highly dynamic conditions. The model is systematically evaluated on multidimensional cloud resource usage data from perspectives such as hyperparameter sensitivity, environmental perturbation adaptability, and data distribution variation. Experimental results show that the proposed method outperforms several existing models in key metrics, including AUC-ROC, AUPR, F1-score, and detection latency, achieving high-accuracy and low-latency anomaly detection in complex dependency environments. This framework provides a scalable, interpretable, and efficient technical pathway for intelligent operation and dynamic dependency modeling in cloud service systems, offering valuable insights for the field of multidimensional time-series anomaly detection.