A University Student Dropout Detector based on Academic Data - A case study at FPT University

Abstract

Dropout at university has become a controversial problem in recent years since the crisis caused many severe consequences for students and universities. FPT University's (Hoa Lac campus) reputation and finances are also affected by student dropout. Therefore, we carried out our research on the early dropout prediction problem to provide school administrators with warning about students who have the risk of dropout so that the school can give proper solutions and support to those students. Our thesis is based on academic performance’s influence on student dropout status. With FPT University, which includes information about students, subjects, and academic performance, we create a dataset that extracts features from the raw database to summarize critical information and partition features with similar characteristics into groups. In addition, we divide the problem into two phases based on FPT University program structure, which includes English preparation terms and Main terms. While FPT University’s database consists of much valuable and massive information, the data dropout status is imbalanced, and many essential values are missing. With the generated datasets and the advance of deep learning neural networks, our research proposed three deep learning models: the convolution-based model (CNN model), the graph convolution network-based model (GCN model), and the tabular learning model (TabNet). Furthermore, compare the deep learning network with traditional machine learning algorithms: logistic regression (LR), support vector classifier (SVC), and light gradient boosting machine (LGBM) with feature selection supported. As a result, the proposed deep learning network performs better than tree-based algorithms, with 72% balance accuracy in the English preparation phase and 75% balance accuracy in primary terms. While TabNet trades off precision to achieve better recall, CNN and GCN models have more balanced results.