Modeling and Predicting the Active Video-Viewing Time in a Large-Scale E-Learning System

Many studies of the mining of big learning data focus on user access patterns and videoviewing behaviors, while less attention is paid to the active video-viewing time. This paper pinpoints this completely different analysis unit, models the extent to which factors influence it and further predicts when a user permanently leaves a course. The goal is to provide new insights and tutorials regarding data analytics and feature subspace construction to learning analysts, researchers of artificial intelligence in education and data mining communities. To this end, we collect video-viewing data from a large-scale e-learning system and use the Cox proportional hazard function to model the leaving time. The models mainly include the interactions between variables, non-linearity assumption and age segmentation. Finally, we use the collected hazard ratios of model covariates as the learning features and predict which users tend to prematurely and permanently leave a course using efficient machine learning algorithms. The results show that, first the modeling can be used as an efficient feature extraction and selection technology for classification problems and that, second the prediction can effectively identify users' leaving time using only a few variables. Our method is efficient and useful for analyzing massive open online courses.