A Dimensionality Reduction Method for Time Series Analysis of Student Behavior to Predict Dropout in Massive Open Online Courses
Résumé
Student attrition is one of the most frequently stated problems with massive open online courses. Although there is a growing body of research that investigates the factors leading to withdrawing from a course, there is also a need for data-driven solutions for early detection as a means to take remedial action. In this case study, we examine the use of several data preprocessing techniques to model attrition on the basis of students’ interactions with course materials and resources. Data for this study were obtained using the Open University Learning Analytics Dataset (OULAD), enabling the analysis of daily summaries of clickstream data. The data were segmented using a variable-sized overlapping window to take into account assignment submission dates as a context-sensitive factor for student attrition. In each sliding time window, features were extracted to characterize student interactions with curricular materials and converted to a set of linearly uncorrelated principal components. The analysis demonstrates that relatively accurate detection of the likelihood of students dropping out from a course can be attained within approximately 10 weeks from the course start date or before completion of assignments worth 20% of the final grade. Although a decision tree model outperformed alternative approaches to model student attrition, a log-linear model performed comparably well on the sparse representation of student interactions obtained through principal components analysis. We discuss the implications of these findings for early identification of at-risk students and the design of analytics dashboards for advisors and tutors.