A Random Forest-Based Self-training Algorithm for Study Status Prediction at the Program Level: minSemi-RF

Abstract

Educational data mining aims to provide useful knowledge hidden in educational data for better educational decision making support. However, a large set of educational data is not always ready for a data mining task due to the peculiarities of the academic system as well as the data collection time. In our work, we focus on a study status prediction task at the program level where the data are collected and processed once a year in the time frame of the program of interest in an academic credit system. When there are little educational data labeled for the task, the effectiveness of the task might be affected and thus, the task should be considered in a semi-supervised learning process instead of a conventional supervised learning process to exploit a larger set of unlabeled data. In particular, we define a random forest-based self-training algorithm, named minSemi-RF, for the study status prediction task at the program level. The minSemi-RF algorithm is designed as a combination of Tri-training and Self-training styles in such a way that we turn a random forest-based self-training algorithm to be a parameter-free variant of the Tri-training algorithm. This algorithm produces a final classifier that can inherit the advantages of a random forest model. Based on the experimental results from the experiments conducted on the real data sets, our algorithm is proved to be effective and practical for early in-trouble student detection in an academic credit system as compared to some existing semi-supervised learning methods.