AI-Based Prediction of Student Dropout Using an Enhanced Deep Learning Framework: DeepDropNet Model

Abstract

In the field of online and higher education, student dropout remains a significant challenge, as the success rate and reputation of educational institutions are closely linked to student retention. In increasingly competitive academic environments, institutions must understand the factors that lead to student disengagement and withdrawal. Student dropout not only affects the individual learner’s academic journey but also impacts institutional metrics such as graduation rates and funding. Numerous studies have been conducted to predict student dropout; however, evaluations of these methods indicate considerable scope for improvement. Therefore, we propose in this research a new student dropout prediction architecture, DeepDropNet, designed to forecast dropout risk in educational settings. Our proposed DeepDropNet combines a 1D convolutional layer with residual blocks, squeeze-and-excitation blocks, and a spatial attention module to improve predictive accuracy. The residual blocks help address the vanishing gradient problem, while the squeeze-and-excitation and spatial attention components enable the model to capture complex dependencies across temporal and behavioral features. To evaluate the model’s performance, we conduct experiments on two real-time educational datasets. Given the common issue of class imbalance in dropout prediction tasks, we employ two data balancing techniques: SMOTE, SMOTEEN, and SMOTETomek. Using 10-fold cross-validation, we thoroughly assess the model, and the results show that DeepDropNet outperforms existing approaches, achieving 94.68% and 96.41% accuracy across the two datasets used.