Advancing Models Portability of Students’ Failure Prediction Using Ontology Modeling Approach

Abstract

Predicting student failure has recently become a major topic of interest for academic researchers. Research in the field of e-learning focuses mainly on the performance and accuracy of machine learning models designed for specific courses. Within the existing literature, not much attention has been given to the application of a prediction model, initially developed for a specific course, and across various other courses. By “model portability”, we mean the capacity of a machine learning model to be applied in many different courses and platforms while maintaining consistent or minimal loss in accuracy. Several factors can significantly affect the portability of a model. The purpose of this study is to evaluate the portability of models developed from data extracted from Moodle logs, augmented with an ontology layer. The adopted approach aims to determine how the quantity of data extracted from Moodle activity traces, the type of attributes chosen (numerical or discretized), and particularly the integration of an ontological structure, affect the portability of models to predict student performance. We applied the K-NN classification algorithm to a set of courses at a similar level to build a model. Then, we evaluated the transferability to other courses by assessing accuracy. The results show that the portability of machine learning models and their implementation with different courses is possible in some cases with an accepted decrease in accuracy. Moreover, the findings demonstrate that the use of an ontology allows a notable improvement in terms of portability.