A Comparative Machine Learning and Deep Learning Models with ElasticNet Regularization for Predicting Student Outcomes: LGBM, CatBoost, ANN, DNN, and WDNN

Abstract

Predicting student academic outcomes is critical for enhancing personalized learning and enabling timely interventions for at-risk students. This study presents a comprehensive comparative evaluation of machine learning and deep learning models—specifically LightGBM, CatBoost, ANN, DNN, and WDNN—enhanced with ElasticNet and Lasso regularization to address challenges of high-dimensional educational data. Using the Math and xAPI datasets, thirteen AI models were evaluated through holdout and k-fold cross-validation across 100 iterations to assess predictive accuracy, generalizability, and interpretability. Results show that CNN with ElasticNet consistently achieves the highest accuracy (up to 93.67%), while ANN performs optimally with Lasso, demonstrating the effectiveness of regularization in improving model stability and reducing overfitting. The findings also highlight the practical utility of ensemble and deep learning models for early detection of at-risk students and support the development of explainable AI frameworks for educational analytics. By addressing prior research limitations, including narrow dataset scope and the absence of advanced hybrid models, this study advances scalable, interpretable, and reliable predictive systems, aiding institutions in data-driven educational decision-making.