ANALYTICAL STUDY OF MACHINE LEARNING & DEEP LEARNING BASED MODELING OF SYSTEMS USED IN EDUCATION

Abstract

Higher education and the popularization of Learning Management Systems (LMS) have led to the creation of massive amounts of data on interaction with learners, which has opened the opportunities of predicting student outcomes based on data. At-risk students are one of the most important issues that higher educational institutions may have to face because early interventions can dramatically enhance the chances of retaining the vulnerable groups and achieving greater academic outcomes. Conventional monitoring methods based mostly on demographics and periodic tests do not normally provide adequate coverage of the dynamic and changing nature of student behavior in online educational settings. This paper involves a systematic comparative study of the analytical machine learning model and the deep learning model in making predictions of student academic outcomes using real LMS data. In particular, a feedforward deep neural network and a Random Forest classifier were compared to each other in terms of their performance under the same experimental conditions with the use of the Open University Learning Analytics Dataset (OULAD). The binary classification problem was formulated in the prediction task to determine successful and at-risk students. Accuracy, precision, recall, F1-score, and ROC-AUC were used to measure model performance, but specific focus was on recall as one of the indicators of early risk detection. Experimental findings prove that although the Random Forest model had good and consistent performance (accuracy rate of 91%), the deep learning model greatly outperformed the random forest in all the metrics, with an accuracy rate of 96.83 and a recall rate of 96.93. In order to overcome the interpretability constraint of the complex models, SHAP-based Explainable Artificial Intelligence (XAI) methods were used on the two models, which explained that assessment performance, LMS frequency of interactions, assignment submission behavior, and forum participation were the most influential predictors. The results indicate the trade-off between predictive accuracy and interpretability and that the integration of deep learning with explainable AI methods can result in trustworthy, clear and practical learning analytics systems to identify vulnerable students in higher education at the first stage.