Understandable Prediction Models of Student Performance Using an Attribute Dictionary

Abstract

This paper proposes a new approach for predicting final student grade with high accuracy. It builds an attribute dictionary (AD) automatically from students’ comments collected after every lesson. Furthermore, it combines white-box models: Decision Tree (DT) and Random Forest (RF), and a black-box model: Support Vector Machine (SVM) to construct an interpretable prediction model and carry out eclectic rule-extraction. First, the AD is built from students’ comments, which are converted to attribute vectors. Second, the output decision is generated by SVM using the attribute vectors in the training phase and then DT and RF are applied to the output decision to extract symbolic rules. Experimental results illustrate the validity of the AD constructed automatically and the superiority of the proposed approach compared to single machine learning techniques: DT, RF and SVM.