A Meta-Learning Framework for Cross-Service Elastic Scaling in Cloud Environments

Abstract

This paper addresses the challenge of limited generalization and adaptability in elastic scaling strategies for service instances under cloud computing environments. It proposes a meta-learning-based framework for cross-service scaling strategy modeling. The framework integrates a service-aware task construction mechanism with a dual-stage strategy prediction model. By extracting transferable knowledge through multi-task learning, the method enables fast adaptation and high-accuracy prediction for new service scenarios. In the task construction phase, the framework introduces service context representation and structural information. This leads to meta-task partitioning with stronger semantic consistency, improving both learning stability and model generalization. In the strategy prediction phase, a dual-stage model architecture is designed. It combines meta-initialized parameters with local fine-tuning. By fusing global coarse prediction with local refinement, the model generates scaling strategies that balance global knowledge transfer with service-specific modeling. Experiments are conducted on a real-world cloud service dataset. The model is systematically evaluated across multiple dimensions, including accuracy, robustness, and adaptability. Results show that the proposed method outperforms mainstream approaches across key metrics and demonstrates strong transferability. In addition, ablation studies and sensitivity analyses confirm the individual contributions of each module to strategy performance. These findings highlight the effectiveness and practicality of the method in complex service scheduling scenarios.