Propuesta y aplicación de métodos de aprendizaje automático para el análisis de datos de educación superior y predicción del éxito académico

Abstract

CONDITION FOR PUBLICATION OF PROJECT. Predicting academic success in higher education is essential for supporting student retention and informing institutional decision-making. This study compares four supervised machine learning models—Logistic Regression, Support Vector Machines, Random Forest, and Multilayer Perceptrons—to predict the performance of first-semester students using academic variables, emotional intelligence indicators, family functionality measures, and behavioral factors. Two criteria of academic success were examined: a strict definition (Y1) and a more flexible one (Y2). The methodology involved data preprocessing, an iterative hyperparameter tuning strategy, and the design of two weighted polynomial kernels that incorporate Logistic Regression coefficients to enhance interpretability. Under the strict criterion, the customized kernel KN1 achieved the highest recall, whereas under the flexible criterion, KN2, Random Forest, and MLP achieved the best performance. Across models, prior academic achievement was the most influential predictor, followed by the course load in the admission program, emotional clarity, and other emotional and family-related variables with moderate but stable effects. The findings indicate that interpretable configurations can achieve competitive performance compared with more complex architectures, offering practical value for institutions seeking to implement early warning systems and develop policies that support students during their transition to university life. Keywords: machine learning, academic success prediction, interpretability, educational data mining, SVM.