Characterization and prediction of academis performance in higher education in engineering using machine learning techniques

Abstract

Today, academic performance in higher education engineering programs remains a key indicator of student retention and educational quality. The early identification of at-risk students poses an institutional challenge for universities due to the multiple factors involved in each case. The main objective of this study is to compare the predictive power of two categories of information: (1) early academic variables and (2) socioeconomic variables. The methodology employed was a quantitative approach based on educational data mining techniques. Thus, predictive models were built using Random Forest with cost-sensitive classification, which made it possible to increase sensitivity in identifying at-risk students. In addition, the relative importance of the variables was quantified, and clustering techniques were applied to analyze the structural coherence of the institutionally defined target variable. The results show that first-year academic performance has greater discriminatory power than the socioeconomic variables considered in isolation. Furthermore, performance in basic science subjects was identified as the most relevant structural predictor. Consequently, this study has shown that early academic data constitute a solid and reliable source of information for identifying academic risk in engineering. Finally, this study provides empirical evidence for the design of early-warning systems in engineering schools.