A Comparison of Deep Learning Techniques for Predicting Students' Academic Performance

Abstract

Higher educational institutions are undergoing a significant digital transformation, driven by rapid advances in artificial intelligence and the growing availability of educational data. Predicting students’ academic performance has become an important analytical task for higher education institutions, as it enables the early identification of students at risk of academic underperformance and supports timely intervention. This study aims to compare the performance of four deep learning models for predicting students’ academic performance using academic and demographic data, identify the most influential predictive features, and examine performance differences across gender and academic majors. The dataset for study is the academic achievement data of 2,878 students from Faculty of Economics & Commercial Studies during the academic year 2017-2019. Four deep learning models were evaluated (MLP, CNN,LSTM and DNN). The results demonstrated that the LSTM model achieved the highest predictive accuracy of 97.14%, outperforming the other models. Feature-importance analysis further revealed that prior academic achievement, particularly second-semester GPA (GPA2), was the most influential predictor of subsequent academic performance. The findings also indicated that gender had limited influence on predictive performance, whereas differences across academic majors were more pronounced. Overall, this study provides a methodological framework that can support higher education institutions in developing data-driven academic prediction systems for early identification and intervention. Future research is recommended to investigate more advanced deep learning architectures, particularly hybrid CNN–LSTM models, to leverage both local feature patterns and temporal dependencies in sequential academic data, potentially improving predictive accuracy and computational efficiency.