An explainable machine learning and human-computer interaction analytical framework for predicting learning outcomes in public music education in minority regions

Abstract

Public music education in minority regions of Southwest China faces constraints related to educational resources and digital infrastructure. Integrating machine learning (ML) with human-computer interaction (HCI) offers a means of analyzing learning behaviors, perceived platform experience, and learning outcomes. This study develops and evaluates a theory-informed, ML-enhanced HCI analytical framework for predicting learning outcomes and identifying interpretable predictive patterns in public music education. The framework is predictive and diagnostic rather than intervention-based. Data were collected from 1,200 non-music-major students across four universities in Southwest China. XGBoost, Random Forest, Support Vector Machine, and a voting regressor were evaluated for predicting learning outcomes. SHAP was used for model interpretation, while adapted SUS- and NASA-TLX-informed measures assessed perceived HCI experience. Assessment-overlap-controlled modelling, leave-one-university-out validation, and university-level sensitivity analyses were conducted to evaluate predictive robustness and cross-institutional stability. Additional measurement-model analyses were used to assess the reliability, convergent validity, and discriminant validity of the adapted HCI measures. In the principal assessment-overlap-controlled analysis, XGBoost achieved R² = 0.736 for the examination-only outcome. For the composite final-score outcome, XGBoost achieved R² = 0.812 after predictors directly overlapping with assignment and participation components were removed. The full-feature models yielded higher performance, with the voting regressor achieving R² = 0.856 (RMSE = 4.18) and XGBoost achieving R² = 0.847; these higher estimates were interpreted cautiously because of predictor–criterion assessment overlap. For the same examination-only outcome and assessment-overlap-controlled predictor set, leave-one-university-out validation yielded a mean R² = 0.713 ± 0.028 across the four institutions. Adding university fixed effects changed test performance only slightly(R² = 0.742 versus 0.736 without university indicators). Video completion, assignment submission, and satisfaction were the leading factors associated with model predictions. Video completion showed an observed nonlinear region between approximately 50% and 70% in the fitted model rather than a validated intervention threshold. Frustration showed a statistically significant but very small association with performance (r = −0.087, p = 0.003) . A Welch heteroscedastic robustness analysis of ethnic-group final scores remained nonsignificant (F _ Welch(14, 156.8) = 1.37, p = 0.176); this finding does not establish ethnic-group equivalence, algorithmic fairness, or educational equity. The framework provides interpretable predictive evidence and generates evidence-informed platform-design hypotheses. Because the study is observational, it does not demonstrate causal effects, improvements in educational efficiency, or the effectiveness of the proposed design strategies. Its principal contribution lies in integrating multi-source predictive learning analytics, cross-institutional validation, and explainable model interpretation in an under-researched public music-education setting, rather than in demonstrating an implemented HCI intervention. Future research should use direct time-on-task measures and multi-region longitudinal or experimental validation.

The authors' abstract, as published at the source. Scientific Reports, 2026 · DOI ↗