Deep Learning-Based Multi-View Behavioral Modeling and Trend Regression for Cloud Resource Prediction

Abstract

This paper proposes a prediction framework for cloud computing environments that integrates a Behavior-Aware Modeling Module (BAMM) and a Multi-Scale Trend Regression Mechanism (MSTRM) to address the challenge of modeling complex access behaviors and their impact on bandwidth usage. BAMM employs multi-view embedding to jointly represent multi-dimensional features such as request type, access path, tenant information, and geographic location. It combines dilated temporal convolution and self-attention mechanisms for multi-dimensional temporal modeling and uses a service call graph for cross-service structural modeling. These three pathways are fused through residual connections and gated mechanisms to generate behavioral representations that capture both fine-grained local patterns and global structural dependencies. On this basis, MSTRM performs multi-scale regression modeling to simultaneously capture short-term fluctuations and long-term trends in bandwidth usage, enabling accurate prediction across multiple periods. The framework incorporates a task-specific loss function to balance prediction accuracy across different time scales while maintaining stable performance in dynamic multi-tenant scenarios. Experiments on a real-world cloud workload dataset demonstrate that the proposed method offers significant advantages in accurately characterizing bandwidth usage trends, reducing prediction errors, and improving robustness in complex environments, providing strong support for high-precision and fine-grained resource usage prediction.