Global Context Modeling and Structure-Guided Feature Enhancement for Queue Time Prediction in Large-Scale Cloud Systems

Abstract

This paper proposes a waiting time prediction framework that integrates global semantic modeling with structure-aware enhancement, aiming to improve the accuracy and stability of task response time in complex scheduling systems. A Global Context-Aware Regression Module is designed to model deep semantic dependencies between tasks and system states, enhancing the model's understanding of scheduling behavior. In addition, a Structure-Guided Feature Enhancement mechanism is introduced, using scheduling structures and task-phase information as priors to perform residual modeling and semantic alignment of intermediate features. This improves representation consistency and discriminative ability under task concurrency and resource fluctuations. Experiments are conducted on a public scheduling dataset, and multiple evaluation metrics are used to compare prediction performance across models, demonstrating the proposed method's advantages in accuracy, robustness, and structural generalization. Further ablation studies and hyperparameter sensitivity analysis confirm the collaborative effect of each module in maintaining model stability and representation capability, validating the effectiveness and applicability of the overall framework.