Analysis of Machine Learning Models for Academic Performance Prediction

The prevalent issue of increased student dropouts, shared by universities worldwide, often culminates in decreased academic performance and prolonged completion times for degree programs. Prompt detection of those students facing a likely chance of failing a course could allow universities to intervene with sufficient support and guidance, facilitating an improvement in their performances. Numerous studies have explored the problem of performance prediction from various perspectives using different representations, algorithms, and data sets. The diversity in research strategies, however, complicates comparisons. In this study, we present a thorough evaluation of various predictive algorithms, representations, and predictive targets for the task of predicting student performance across 77 different courses in three distinct programs at the Universidad de los Andes: Systems and Computer Engineering, Industrial Engineering, and Economics. The results show that representing data in windows of time spanning 3 previous semesters, in conjunction with the LSTM-based algorithm for binary classification, yields the best results, achieving a precision of 0.838.