Predicting Student Dropout Risk Using Long Short-Term Memory Based on Academic Performance Data

Abstract

Student dropout is a critical issue affecting academic quality and institutional performance in higher education. Dropout behavior usually emerges gradually through declining academic performance across semesters. Therefore, time-series modeling is essential to capture such temporal patterns. This study proposes a Long Short-Term Memory (LSTM) model to predict student dropout risk based on semester-wise academic data. The dataset consists of 385 undergraduate students from the Computer Science program, FMIPA, represented by Grade Point Average (GPA) and credit load (SKS) over eight semesters. Student status is converted into a binary label: dropout and non-dropout. To address class imbalance, Synthetic Minority Over-sampling Technique (SMOTE) is applied. Experimental results show that the proposed LSTM model achieves a recall of 1.00 for the dropout class, indicating that all dropout cases are successfully detected. Although the precision remains relatively low due to false positives, the model demonstrates strong potential as a basis for academic monitoring and early intervention systems.