An Integrated Framework for Fair and Interpretable Predictive Modeling in Educational Analytics

As learning analytics becomes more widely used in education, there is a need to develop predictive models that account for a range of real-world factors. However, existing predictive systems emphasize model performance more than the design of transparent, trustworthy models. This study combines model performance, fairness, and interpretability in a single pipeline. The integrated predictive framework evaluates five classification models under stratified cross-validation on the Open University Learning Analytics Dataset (OULAD). Two data files are selected from the dataset for model development. The processed data comprises 32,593 merged records with demographic and behavioral features of the learners. SVM-RBF and gradient boosting classifiers show similar, high predictive performance in mean Accuracy, AUC, and F1-score. Fairness is assessed using gender as the sensitive attribute. Random forest shows the lowest group disparities, with a mean Demographic Parity Difference (DPD) of 0.1147 and an Equalized Odds Difference (EOD) of 0.1610. Statistical analysis further identifies significant classifier differences in all metrics except DPD; however, pairwise tests show no significant difference for any metric. SHAP provides feature importance plots for model interpretation based on the sensitive attribute. They show that the behavioral features contribute more to predictions than demographics and primarily drive model decisions. Finally, the study offers practical guidance for developing transparent and fair educational AI systems.