Exploratory machine-learning classification of student performance in physical education using fitness, engagement, and academic indicators

class imbalance data leakage exploratory classification machine learning physical education student performance

Authors

Articles in Press
Original Research

Background: Physical Education (PE) performance is multidimensional and may be associated with fitness, engagement, attendance, motivation, and prior achievement. Machine learning can identify classification patterns across these indicators, but predictive claims require clear temporal ordering and rigorous validation.

Objective: This study compared four machine-learning classifiers for categorizing student PE performance while minimizing target leakage and addressing class imbalance.

Methods: An exploratory, leakage-controlled classification analysis used a public Kaggle dataset of uncertain provenance. The dataset contained 500 records, 19 substantive variables, and one identifier, with Performance categorized as Low, Average, or High Performer. Because the label-construction rule was undocumented, Performance was treated as an inherited categorical outcome. Overall PE Performance Score, Improvement Rate, and Final Grade were excluded because they could contain current, change-based, or outcome-related information. Categorical predictors were one-hot encoded, numeric predictors were standardized, and preprocessing was performed within training folds. Random Forest, Decision Tree, Support Vector Machine, and Logistic Regression were evaluated using stratified five-fold cross-validation with class-weighted learning.

Results: Random Forest achieved the highest accuracy (0.856), whereas Logistic Regression achieved the highest balanced accuracy (0.749) and macro-F1 (0.645). Its out-of-fold recall was 0.759 for Low, 0.720 for Average, and 0.750 for High Performers, indicating that accuracy alone may inadequately reflect performance in an imbalanced multiclass setting.

Conclusion: Leakage-controlled, class-aware evaluation is essential in machine-learning classification of student performance. These findings are exploratory and should not be interpreted as validated prospective predictions. Future studies should use documented data, explicit temporal ordering, larger balanced samples, and independent external validation.