Modelos complexos de predição aplicados na educação

Abstract

The current doctoral thesis presents a compilation of five papers employing complex predictive models to solve educational research issues. The first paper presents the classification and regression trees, as well as bagging, random forest and boosting algorithms. They are used to create an academic achievement predictive system, using a set of cognitive assessments/tests as independent variables (or predictors). The second paper, by its turn, uses the random forest algorithm to predict the academic achievement of high-school students. Once again, a set of cognitive assessments/tests were used as predictors. In the third paper, we introduce a new visualization technique that enables to visually inspect the quality of the prediction made using random forest. This technique is based on the plot of statistical information as a weighted graph, enabling the use of additional prediction quality indexes beyond total accuracy, sensitivity and specificity. The fourth paper presents the random forest algorithm as an imputation method, and investigates its impact on item fit to the dichotomous Rasch model and on their difficulty estimate. Finally, the fifth paper compares the classification tree with a Naïve Bayes classifier in the prediction of academic drop-out, using a set of socio-demographic variables as predictors. The papers presented in this doctoral dissertation introduce a set of innovative quantitative methods that have potential to solve a number of issues in the educational research field. They can also led to new discoveries, not allowed by other, more classical, methods.