MPC-SDP: um classificador genérico baseado em múltiplas perspectivas para a predição da evasão estudantil

Abstract

Given the severe issue of student dropout, many researchers have employed data mining techniques on large volumes of educational data to develop models that identify at-risk students and factors. The studies usually represent students’ behavior based on their academic, social, and economic aspects. Developing models specialized in degrees and prediction moments was common, although their results needed to be more comprehensive. Additionally, studies generally apply the techniques in a traditional way without adapting them to the specificities of the dropout domain. In order to address these gaps, this thesis proposes the development of the MPC-SDP, an ensemble model designed to predict dropout in different degrees and stages of the academic trajectory (generic prediction) but which is guided by specialized perspectives that are important for the domain of student dropout. More specifically, the MPC-SDP aims to (i) give more attention to the variability of dropout patterns along the students’ academic trajectory through the construction of an ensemble made up of sub-classifiers specialized by academic periods/semesters and (ii) enable broader participation of different aspects of student data (academic, contextual, economic, interactional, and social) in representing dropout behavior, via an aspect-oriented attribute selection process. When evaluated and compared to generic prediction models built traditionally, the MPC-SDP improved the precision of predicting students who drop out in the initial semesters, considering the context of face-to-face undergraduate courses at the Instituto Federal Farroupilha as a case study. This feature would better target preventive actions at the most critical stages of student follow-up since the initial semesters have the highest enrollment and dropout figures. Additionally, the aspect-oriented feature selection increased the variability of aspects represented among the main patterns of the MPC-SDP sub-classifiers, potentially contributing to a broader and multifaceted understanding of their dropout patterns.