A New Lens for Looking at MOOC Data to Predict Student Performance

Abstract

We present two novel methods to predict students' grades using their action time series in Massive Open Online Courses (MOOCs). The main motivation behind this contribution comes from three main differences in the methods used in previous research. First, the methods used to analyze time series often aggregate the data, which discards the effect of the previous actions on the present actions. Second, most of the previous research has a common assumption that the actions are distributed homogeneously in time, which might or might not be true for students in MOOCs. Third, the methods used to predict students' grades are often based on linear regressions and correlations, which assume a normal distribution for the data generation processes, which might not be true in all cases. To highlight the first two differences we propose to use Generalized Auto Regressive Conditional Heteroskedasticity (GARCH) models. To deal with the third difference, we propose to use the Extreme Values Theory (EVT). The results show a significant improvement over existing methods in terms of prediction abilities. The results also show that there is an improvement even if a shorter time series data was used.