Adaptive Weighted Ensemble Learning with Hybrid Feature Selection for Explainable Student Performance Prediction
Abstract
Student identification of educationally at-risk students is a critical component of the effectiveness of learning outcomes, student retention and timely interventions. But making a correct prediction of student performance is still difficult because of educational data that is high-dimensional, student learning behaviors that vary and the defined interpretability of many machine learning models. In this paper, an Adaptive Hybrid Feature Selection and Weighted Federated Explainable Learning (AWFA-LightGBM) framework is proposed to feature selection and prediction early student performance in Open University Learning Analytics Dataset (OULAD). A hybrid feature selection methodology based on Variance Threshold (VT), Mutual Information (MI) and Boruta is exists used initially to pick the highly informational features from the original feature space and nine salient features are selected. Random Forest, XGBoost, and LightGBM are domesticated in isolation and then the prediction probability of each model is adaptively combined with a Weighted Federated Aggregation (AWFA) mechanism, based on the validation performance. Moreover, a set of interpretable and transparent feature-level insights of the prediction process is given through the use of SHapley Additive exPlanations (SHAP). Experimental results prove so the proposed AWFA-LightGBM framework surpasses the performance of individual miniatures such as Random Forest, XGBoost and LightGBM with an accuracy of 94.02%, a precision of 91.25%, a recall of 96.59%, an F1-score of 93.84%, ROC-AUC of 98.32%, Matthews Correlation Coefficient (MCC) of 88.17% and Cohen's Kappa of 88.03%. In addition, the total score, average assessment weight, assessment count, and average score are the most significant factors to anticipate student academic performance in the SHAP analysis. The proposed system offers a reliable, interpretable and accurate decision-support framework to predict student performance in higher education.