Evaluation of Prediction Algorithms in the Student Dropout Problem

Abstract

University dropout is a growing problem which, in recent years, is using computer techniques to assist in the detection process. The paper presents the evaluation of some prediction algorithms to detect a student with a high possibility of scholar desertion. The approach uses real data from past scholar periods to create a dataset with different information of the students (i.e., personal, economic, and academic records). The algorithms selected in the experimental phase were: J48 decision tree, K-near neighbors, and support vector machine. We use two similarity metrics to split the dataset with cases with at least 80% of similarity to evaluate each case. We use the data from 2010 to 2016 with real students’ information to predict if there exists the possibility of a real academic dropout in one test for a period. The results show that the J48 algorithm reaches a better performance in both experiments. Besides, the tree generated for each student is taken as a path of attention, reaching around 88% of effectiveness. Finally, the conclusions argue the contributions of the paper and propose a future line of research.