Machine Learning and Predictive Learning Analytics in Massive Open Online Courses

Abstract

In this paper, we develop predictive learning analytics for online classes to predict student performance on in-video quizzes in Massive Open Online Courses (MOOCs). We parse clickstream data from a sample MOOC to create 9 features (including “Quiz duration” and “Number of rate changes”) that summarize learner behavior for each user-video pair. We also use the clickstream data to create 2 sets of position-based features- one regarding the video segments the user visits and the other regarding the transitions that the user makes. We create a variety of machine learning models that use these features to predict student performance on quizzes. This performance is modeled as a binary value, where “1” indicates that the learner was correct and “0” shows that they were incorrect. Our best performing model created with only behavioral features, a neural network, has an average accuracy of almost 70% which represents an improvement of 3.27% over the baseline. On the other hand, the neural network created with only positional features has an average accuracy that matches the baseline. However, the neural network we made with both the position and behavioral features is our best performing model with an improvement of 3.75% over the baseline. Overall, these results show that there is potential for student behavior to predict performance in MOOCs. In future research, we will continue to modify and add to our features and models to improve their performance to a level that would help enhance the online learning experience for both students and instructors.