An University Student Dropout Detector Based on Academic Data

Dropout at university has become a controversial problem in recent years, resulting in severe consequences for students and universities. Our research focuses on predicting early dropout to provide school administrators with warnings about at-risk students. This enables schools to offer appropriate solutions and support. This research explores the influence of academic performance on student dropout. We evaluated the proposed method on a dataset using comprehensive information about students, subjects, and academic performance. By extracting features and grouping them based on similar characteristics, we summarized critical information from the raw database. The problem is divided into two phases: English preparation terms and Main terms, aligning with the university's program structure. However, the data on dropout status is imbalanced, and numerous essential values are missing. To address this, we employed three deep learning models: the convolution-based model (CNN), the graph convolution network-based model (GCN), and the tabular learning model (TabNet). We compared the performance of these deep learning models with traditional machine learning algorithms: logistic regression (LR), support vector classifier (SVC), and light gradient boosting machine (LGBM) with feature selection. Our deep learning models outperformed tree-based algorithms, achieving a balanced accuracy of 72% in the English preparation phase and 75% in primary terms. It is worth noting that TabNet prioritizes recall at the expense of precision, while CNN and GCN models yield more balanced results.