A Machine Learning Approach for Web Service Autoscaling: EMD-LSTM and Attention Mechanism in Cloud Computing

Traditional reactive autoscalers respond only after resource exhaustion occurs, causing SLA violations and oscillating scaling behaviour. This paper proposes a proactive autoscaling system that combines Empirical Mode Decomposition (EMD), LSTM networks, and attention mechanisms to predict web service traffic and provision resources. The approach decomposes non-stationary RPS patterns into 7 Intrinsic Mode Functions, predicts each component using independent LSTM models, and weights predictions using multi-head attention to emphasise informative signals. Evaluation on real e-commerce traffic demonstrates that EMD-LSTM achieves an R² of 0.991, compared with 0.223 for the baseline LSTM. Production deployment in Kubernetes environments shows the system processed 7.67× more requests while consuming only 72.7% fewer resources, achieving a 28x improvement in resource efficiency. The work shows that combining accurate signal decomposition, temporal learning, and attention-based weighting simultaneously achieves both SLA compliance and cost efficiency, while reducing the lost request ratio from 87.17% to 1.63%.