Predicting and Interpreting Student Performance in Higher Education: A BOHB-Optimized DNN with Dual Explanation Framework Using SHAP and HFSF on OULAD

This study focuses on predicting student performance and identifying influential factors to optimize educational strategies, utilizing the Open University Learning Analytics Dataset (OULAD). After preprocessing the data, multiple machine learning models including KNN, Logistic Regression, Naive Bayes, SVM, Decision Trees, and Random Forests were evaluated. The core of our approach is an optimized Deep Neural Network (DNN) that employs Bayesian Optimization and HyperBand (BOHB) for hyperparameter tuning, significantly enhancing predictive accuracy. For robust feature analysis, a novel Hybrid Feature Scoring Function (HFSF) and an interpretable SHAP analysis were implemented on the optimized DNN to rank feature importance. The results demonstrate that the proposed BOHB-optimized DNN outperforms all baseline models in categorizing students into distinction, pass, fail, and dropout classes, while also identifying key influencing factors such as demographics, prior education, and online resource engagement. This work contributes an effective predictive framework for educational systems and highlights the potential of hybrid optimization techniques, with future research directed toward hybrid models and longitudinal data for improved accuracy and generalizability.