LLM-Guided Hierarchical Ensemble for Multimodal Cloud Performance Prediction

Cloud infrastructure performance prediction is important for system efficiency. It is hard because the data comes from different sources like configurations, metrics, and logs. Old methods cannot capture the relationships among these types of data. This paper presents HELM-ECS, a new framework to solve this problem through a unified learning approach. It uses DeepSeek LLM to get features and connect different data. The LLM also helps in guiding how diverse data types are fused at a semantic level, ensuring consistency and preserving their individual meanings. HELM-ECS introduces anomaly-aware decomposition to deal with unstable time series and track transient performance shifts. It adds a confidence-based feature selection method that improves the selection of predictive signals. The system further applies a hierarchical XGBoost ensemble to model complex feature interactions and reduce overfitting. These designs together make the framework strong in handling noisy, dynamic cloud data. HELM-ECS is compact, fast, and scalable, making it suitable for practical cloud monitoring and management systems.