FEAST: FEATURE ENGINEERING AND STACKING FRAMEWORK FOR COMPLEX SYSTEM PREDICTION

The challenge of capturing high-dimensional interactions and ensuring stable predictions in complex engineering systems, characterized by rapidly growing heterogeneous data, limits traditional models. To address this, we introduce FEAST (Feature Engineering and Advanced Stacking Framework), whose core innovation lies in its integrated approach: (1) Hierarchical feature optimization using K-Means++ clustering and adaptive correlation to significantly reduce redundancy and enhance feature discrimination; (2) An advanced stacked ensemble leveraging diverse base learners (MLR, XGBoost, LightGBM, CatBoost, GBDT) fused via Shapley-weighted combination and a regularized (Elastic Net) meta-learner to robustly capture complex patterns; (3) An efficient Monte Carlo Dropout-based uncertainty quantification module providing reliable confidence intervals for risk-aware decisions. Comprehensive experiments demonstrate FEAST's superiority: it achieves 12.7% higher prediction accuracy (p