Assessing the Influence of Time on Features for the Prediction of User Dropout

Parisa Shayan, Martin Atzmueller, Menno van Zaanen

Abstract: This article investigates the influence of time on features for the prediction of user dropout in an online training platform. Specifically, we target a comparison between the two time measurements: activity-based versus duration-based. Considering time, we utilize features in different groups: either activity-based features for activities or duration-based features for duration, as well as the start-based, action-type-based, and course-based features for both time measurements. The most surprising aspect of the results is the high accuracy of the classifiers from the tenth activity (which corresponds to almost half a day on average) onward. While the action-type-based and the course-based features have a major influence on dropout, the start-based features are only influential in the classifiers that use information of activities at beginning. In addition, the activity-based features have only a minor impact in the middle of the course whilst the duration-based features have a major influence throughout the course.