Latent Factors for Consistently Predicting Student Success

Developing Machine Learning models to predict the likelihood of a student's graduation has received significant interest in recent times. One clear application is to identify students most in need of support before actual failure occurs. There is a growing concern, however, about the range of applicability of such models. Machine Learning models are often limited by the consistency of their performance across years or even by programme in other words, although a model may be developed for a given course/module in a given year, the model accuracy tends to degrade when small differences occur in the time or field of study. In this study, the focus is on the identification of so-called Latent Factors, which are more fundamental characteristics derived from the student and field of study meta-data. Basing Machine Learning models on these more fundamental characteristics tends to produce models which, although reduces in accuracy, tend to preserve the prediction capacity over a broader period of time and scale of study area. The study investigates latent factors that include a student's “credit load capacity”, level of activity in accessing course material (LMS access frequency), overall performance (measured as mean marks), the rate of change of performance (measured as the rate of change of mean) and consistency (measured as standard deviation). In addition, the modelling also considers the matric mean score of the students undertaking the coursework, historical consistency with peer modules (given by the Pearson R-Coefficient), course position in curriculum (given by the academic year of study when undertaken by students) and the mean number of attempts required to pass the course. It is shown that when these characteristics are integrated into a Machine Learning framework, the accuracy improves on the order of 24%