Dimensionality Reduction for Predicting Student Performance in Unbalanced Data Sets

摘要

In this study, we evaluated two data sets from two Portuguese schools for predicting student performance. These data sets contain not only the previous grades of the students, but also the demographic, social and school related features. Both data sets are unbalanced in class distribution and contained some irrelevant features. Such characteristics may cause unsatisfactory True Positive (TP) rates for the Fail grade. This grade is important in prediction but it has a low representation as compared with the Pass grade. To improve prediction, dimensionality reduction was performed on both data sets to generate subsets that contained: (i) features selected by a wrapper approach, and (ii) only previous grade(s). The results showed that dimensionality reduction helped to improve the TP rates for the Fail grade. In addition, good classification accuracies were attained. We also noticed that even though the subsets contain only one previous grade, comparable accuracies can also be achieved.