Multi-source feature-based prediction of academic rank among physical education teachers in higher education: A comparative study of XGBoost and BP neural networks

Abstract

As data-driven approaches gain traction in physical education, universities require scientific, transparent, and interpretable tools to support academic-rank decisions. This study addresses limitations of CV-based evaluations by developing a competence-aware prediction framework that integrates multiple sources of information, including teachers’ background characteristics, teaching behavior logs, and self-assessed competencies. Using data on 1,294 full-time physical education teachers from several universities in the 2023–2024 academic year, we model academic rank as a binary outcome that distinguishes intermediate from senior levels. The feature set includes academic degree, teaching experience, annual teaching workload, platform activity indicators, participation in teaching-reform projects, and four competency constructs: teaching design, teaching reflection, technology integration, and classroom management. Two machine-learning models, XGBoost and a backpropagation neural network, are trained and evaluated with stratified five-fold cross-validation and learningcurve analysis across increasing training sizes. Model interpretability is assessed with SHAP (Shapley Additive Explanations). XGBoost outperforms the neural network on the held-out test set across precision, recall, F1 score, and accuracy, achieving an F1 score of 0.96 compared with 0.90, and it shows lower variance across cross-validation folds. Learning curves indicate performance saturation for XGBoost at approximately 500 samples, whereas the neural network remains more sample-dependent. SHAP highlights teaching reflection competence, academic degree, and annual teaching workload as the most influential predictors, followed by technology-integration and instructional-design competencies. These findings indicate that competence- and engagement-oriented features provide stronger and more actionable signals than raw platform activity counts, and that an interpretable ensemble model offers superior risk control by reducing both false negatives and false positives. Integrating behavioral and competency indicators within an interpretable framework can enhance the fairness, transparency, and operational reliability of faculty-promotion processes in physical education. The approach offers practical guidance for evidence-based policy design and can be extended to multimodal evidence, broader institutional contexts, and systematic fairness auditing.