PREDICTING STUDENT ACADEMIC PERFORMANCE: A STUDY OF HYBRID AND ENSEMBLE REGRESSION MODELS ON HABIT-BASED

Abstract

This work examines the extent in which a variety of regression models can be used to predict the academic performance of students based on an enhanced data set of academic habits. The eight models are: classical linear regression, Support Vector Regression (SVR), and ensemble methods (Random Forest, Bagging, and Increasing), as well as the hybrid models combining any two or more of the basic ones. Every model is tested according to a few performance measures with special focus on the overall accuracy and explanatory power. It has been found that the Among the models tested, the Boosting algorithm outperforms others, achieving the best results with an MAE of 3.1947, MSE of 16.8005, RMSE of 4.0988, and an R² of 0.8664. The addition of disparate strengths usually looks appealing in theory since it is possible to see how the skills of each model can be combined in ways that offer sizable performance improvements, but can perform below component models when there is no tuning in the composition of diverse strengths. Therefore, linear regression still serves as quite solid basis, which, at once, manifests evidently the linear elements hidden within the data. Such results raise a concern of careful choice of models and care in controlling complexity when implementing machine learning in education. Also, the combination of hybrid architectures and adaptive weighting schemes could be explored in case one would like to achieve both robustness and interpretability in the process of academic performance prediction.