A novel hierarchical evolutionary feature selection (HEFS) framework is proposed to enhance interpretability and predictive performance in academic performance analysis. The framework addresses the challenge of selecting meaningful features from high-dimensional lifestyle data while preserving psychological validity. Initially, features are organized into semantically coherent clusters through hierarchical agglomerative clustering, employing correlation-based distance metrics and Ward’s linkage to ensure logical groupings. These clusters are subsequently optimized using an evolutionary algorithm (EA) designed to balance predictive accuracy with interpretability, guided by established behavioral theories. The selected feature subsets are further validated against recognized psychological constructs to ensure theoretical relevance, thereby integrating data-driven insights with domain knowledge. The refined feature set is then incorporated into interpretable machine learning models to support transparent and accountable decision-making. Experimental results indicate that HEFS outperforms conventional feature selection methods in terms of both predictive performance and explainability. This study contributes a principled and scalable approach to feature selection that reconciles statistical rigor with psychological grounding, offering practical value for educational analytics and related domains. While the primary validation uses a single-institution dataset, cross-program generalization experiments (AUC-ROC = 0.841 ± 0.004) indicate robustness across diverse curricular structures, though multi-institutional replication remains necessary.
原文链接
分享
引用
