Evaluation of grade prediction using model-based collaborative filtering methods
Journal
2018 IEEE Global Engineering Education Conference (EDUCON)
Date Issued
2018-04
Author(s)
Rechkoski, Ljupcho
DOI
10.1109/educon.2018.8363352
Abstract
Estimating grades for courses that are yet to be enrolled by students can help them in making decisions towards timely graduation and achieving better overall results. This paper presents an evaluation of grade prediction for future courses using the model-based collaborative filtering methods: Probabilistic Matrix Factorization and Bayesian Probabilistic Matrix Factorization using Markov Chain Monte Carlo. The prediction model was evaluated in a simulated scenario of an enrollment cycle in a winter and summer semester, based on a real data-set of enrollments and grades over several years at the authors' institution. Several evaluation metrics were used in order to assess the accuracy of predictions and analyze the distribution of the prediction deviation across study programs and grades. Beside the standard approach in predicting the final grade that is to be achieved by a student in a future course, we have also devised a method to estimate if the student will fail the course, so that he will have to re-enroll it at least once. The results showed that the predicted grades were in the range ±1 compared to the actual grades in more than 80% of the records.