A Behavioural Analytics Framework for Early Performance Prediction in Introductory Engineering Programming Courses

Resumen

Introductory computer science engineering programming courses are associated with high failure and attrition rates; hence they need reliable early student performance prediction to enable timely academic interventions. This study proposes a behavioural analytics-based framework for predicting student performance. The study used over one lakh programming submission records collected from an online programming environment from which total 4,039 student level aggregated activity logs were constructed. Authors then applied advanced feature engineering to these aggregated data to obtain temporal engagement patterns, error rates, effort efficiency, and aggregated engagement measures. Logistic Regression (LR), Random Forest (RF), Gradient Boosting Classifier (GBC), Multilayer Perceptron (MLP), Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost) were tested on an 80:20 train-test split and 10-fold GroupKFold cross-validation. The Authors subsequently developed a stacking ensemble model combining RF, GBC, MLP, SVM, and XGBoost as base learners, with LR as the meta-learner. The ensemble obtained F1-score of 0.857 and ROC-AUC of 0.905 and was more reliable than individual classifiers, Statistical significance of the performance gains achieved through behavioural feature engineering. Authors confirmed Ensemble learning using a paired Wilcoxon signed-rank test. The proposed framework is applicable to detect at-risk students, promotes certain pedagogical interventions in computer science education.