Machine Learning-Based Classification of Student Adaptability in Online Learning with Feature Engineering

Student adaptability in online learning environments has become increasingly important in contemporary education. This study introduces a feature engineering approach guided by SHAP (SHapley Additive exPlanations) to enhance the classification of student adaptability levels. Unlike prior studies that primarily utilize exploratory analysis or statistical importance scores, this method leverages SHAP values to construct new features by considering both statistical contribution and semantic meaning. Three additional features were created by combining original variables, representing educational level and session duration, digital access quality, and socioeconomic context. The dataset was evaluated using four classic machine learning models, namely Support Vector Machine (SVM), k-Nearest Neighbors (KNN), Decision Tree, and Random Forest, both before and after applying the engineered features. Results show that SHAP-based feature engineering improved model performance in most cases. The most notable gains were observed in Decision Tree and Random Forest models, where the F1-score increased from 84.87% to 89.34% and from 85.80% to 89.34%, respectively, while accuracy rose from 88.38% to 90.08% and from 89.63% to 90.08%, respectively. The SVM model also recorded an increase in recall from 82.49% to 87.28%, whereas KNN showed a slight drop in accuracy but improved in ROC AUC from 91.55% to 93.83%. These findings demonstrate that explainable feature design not only enhances accuracy and F1-score, particularly in tree-based models, but also supports model interpretability, enabling more transparent, reliable, and effective educational decision-making systems.