Learning Analytics e Mineração de Dados Educacionais da Teoria à Prática: Aspectos Envolvidos na Implementação em Diferentes Contextos e Níveis Educacionais

Abstract

This thesis seeks to present the similarities and differences of the practical ap plication of Learning Analitycs (LA) and Educational Data Mining (EDM) in different educational contexts and levels. In this way, the theoretical and practical concepts involved in the process of transforming educational data into information and knowl edge in different contexts and levels of education are presented. To accomplish such objective this thesis explains the theories involved, as well as the state of art, methodologies, processes, methods and the context of the applications. Therefore, three distinct use cases and practical applications developed are presented for the following scenarios: face-to-face secondary education in Uruguay, university education in Uruguay and technical secondary distance education in Brazil. In this context, we seek to generate practical methodologies for exploring data from different educational databases and contexts, focusing on generating early warning models for school dropout. The first methodology aims to identify secondary school students at risk of dropout and failure. For this purpose, the models used data collected from 258,440 students in nine different databases. These data were transformed into temporal data and then used to train the classifiers. This application presented good results and is currently under implementation in Uruguay. The second application describes the use of Data Science and EDM techniques with data from 4,529 onsite students at the University of the Republic of Uruguay (Udelar). The main contribution of this approach was the combination of different data sources that demonstrated high predictive power, achieving predictive rates with excellent discrimination in the fourth week of a course. In addition, the analysis showed that students with more interactions within the Virtual Learning Environment (VLE) tend to be more successful in their subjects. Results also revealed some relevant attributes that influenced students’ success, such as: the number of subjects in which the student is enrolled and the mother’s education. From these results, some institutional policies emerged, such as allocating resources for the VLE infrastructure and developing tools for monitoring students. The third approach is a genetic algorithm (GA) that seeks to improve the selection of classifier hyperparame ters. The proposal of this algorithm seeks to increase the accuracy rates of automated models to detect at-risk students enrolled in a technical distance education course in Brazil. The main scientific contribution is the performance of the approach compared to traditional techniques, where GA proved to be a viable alternative, producing better results than conventional techniques in terms of precision and computational cost. This thesis presents the results obtained in these applications, the similarities and differences among them according 19 technical aspects (existence of temporality in the data, size of the databases, existence of multiple data sources, techniques used, among others). Lastly, the thesis establishes the scientific contribution and future work related to the topic.