Early Prediction of University Student Dropout Using Machine Learning Models

Quispe J.O.Q.; Toledo O.C.; Toledo M.C.; Llatasi E.E.C.; Saira E.M.R. · Nanotechnology Perceptions (2024) · DOI: 10.62441/nano-ntp.v20is5.62

Resumen

The study analyzed and validated the best machine learning model to predict dropout among students at the National University of Moquegua, Mariscal Nieto campus, between 2009 and 2019. It also identified the characteristics influencing dropout. The models used were logistic regression, decision trees, support vector machines, naive Bayes, the CRISP-DM methodology, and metrics such as the confusion matrix and K-fold cross-validation. We selected important features using two techniques, random forests and Featurewiz, employing Python and its libraries for the task. The research type is descriptive correlational, and the design is observational using secondary data sources. The sample was selected through stratified random probability sampling, resulting in a sample of 109 students who dropped out without completing their studies and 220 who did not, totalling 329 data points. Cross-validation indicated that decision trees achieved 76% accuracy, followed by logistic regression at 73%, support vector machines at 71%, and naive Bayes at 62%. The characteristics influencing dropout include general data such as cycle, age, and zonal address; economic aspects such as total income, family load, children in higher education, household components, student's housing, and head of the household; and housing aspects such as type of construction and number of bedrooms.

Instituciones: Universidad Nacional del Altiplano; Universidad Nacional Micaela Bastidas de Apurímac; Universidad Nacional de Moquegua