Undergraduate Academic Performance Prediction While Maintaining Both Accuracy And Interpretability

Abstract

Abstract Student dropout in higher education is one of the significant problems encountered by educational institutions and students globally. This theses focuses on identifying contributing factors and improving prediction accuracy using various machine learning techniques. The research uses Logistic Regression, Decision Trees, Random Forest Trees, Support Vector Machines (SVM), Naive Bayes, and Boosting Classifiers like XG Boost Classifier, Gradient Boost Classifier, CatBoost Classifier, and AdaBoost Classifier to examine both academic and non-academic factors. To enhance the analysis, the research incorporates techniques like correlated feature management and hyperparameter tuning, alongside data sampling methodologies such as SMOTE, SVM-SMOTE, and ADASYN with ADASYN emerging as the best sampling technique. After the initial stages the research found that the CatBoost classifier, enhanced by ADASYN sampling, significantly improved prediction accuracy with a testing F1-Score of 0.8603, suggesting a robust model for educational institutions to early identify at-risk students. Then, the second phase of this research was focused on the interpretation of the prediction. There we considered LIME,SHAP and Explainable Boosting Machine as the interpretable and explainable models. This thesis identifies non-academic factors such as socio-economic background and personal resilience as significant predictors of student dropout rates, beyond academic performance alone from the explanations of the two XAI models. This study conducted comprehensive experiments encompassing machine learning and explainable artificial intelligence methodologies, aiming to optimize accuracy and interoperability in the obtained results and found out how LIME and SHAP can be applied for interpretability according to the context.