A Happiness Prediction Model Based on Performance-Diversity Dual-Constraint Dynamic Weighted Ensemble

Abstract

To address the problems of insufficient generalization ability of traditional single machine learning models in happiness prediction tasks and the failure of fixed-weight ensemble models to adapt to sample-level prediction differences, this paper proposes a Performance-Diversity Dual-constraint Dynamic Weighting Ensemble (PD-DWE) model. First, a diversified base model pool including Random Forest, K-Nearest Neighbors, Support Vector Machine, Gradient Boosting Tree, XGBoost, Multilayer Perceptron, and LightGBM is constructed, and hyperparameter optimization of the base models is completed through a two-stage random search. Second, based on the 10-fold stratified cross-validation framework, high-quality model selection is performed using an F1-score threshold in each fold, and diversity constraint is achieved by combining predictive correlation filtering. Finally, a dynamic weight calculation strategy integrating the global performance of models and sample-level prediction confidence is designed, and the optimal classification threshold is adaptively searched to complete the final prediction. Experimental results show that the proposed PD-DWE model achieves an F1-score of 79.1% on the test set, representing a 12.43% relative improvement compared with the optimal single model, and an AUC value of 69.61%. Its comprehensive prediction performance is significantly superior to that of single machine learning models, providing a more stable and efficient technical solution for the quantitative prediction of happiness.