Application and Analysis of XGBoost Models Optimized with TPE in Complex System Prediction

Abstract

This paper proposes an XGBoost model based on TPE hyperparameter optimization for constructing a system effectiveness prediction framework, focusing on integrating precise prediction with interpretable analysis. First, TPE algorithm-based hyperparameter optimization enhances the XGBoost model's predictive performance, elevating the model's decision coefficient from the baseline 0.827 to 0.875. Second, a Monte Carlo simulation combined with confidence interval estimation constructs an uncertainty-quantified prediction system, effectively overcoming limitations of traditional methods to deliver more reliable result intervals. Finally, the SHAP model was applied to analyze feature importance in the XGBoost model, revealing nonlinear relationships between multidimensional features and target responses, thereby enhancing interpretability. This model quantifies the marginal effects of coaching factor inputs through hypothesis testing and develops optimal resource allocation strategies based on reward-punishment balance. The framework significantly improves prediction accuracy, interpretability depth, and strategy feasibility, providing methodological support for complex decision-making scenarios. The innovation of this research lies in integrating XGBoost with the TPE optimization algorithm, which not only enhances the predictive model's performance but also overcomes the limitations of traditional point estimation. It provides a comprehensive perspective encompassing uncertainty quantification and feature analysis, offering broad application value.