Can Feature Predictive Power Generalize? Benchmarking Early Predictors of Student Success across Flipped and Online Courses
Abstract: Early predictors of student success are becoming a key tool in flipped and online courses to ensure that no student is left behind along course activities. However, with an increased interest in this area, it has become hard to keep track of what the state of the art in early success prediction is. Moreover, prior work on early success prediction based on clickstreams has mostly focused on implementing features and models for a specific online course (e.g., a MOOC). It remains therefore under-explored how different features and models enable early predictions, based on the domain, structure, and educational setting of a given course. In this paper, we report the results of a systematic analysis of early success predictors for both flipped and online courses. In the first part, we focus on a specific flipped course. Specifically, we investigate eight feature sets, presented at top-level educational venues over the last few years, and a novel feature set proposed in this paper and tailored to this setting. We benchmark the performance of these feature sets using a RF classifier, and we provide and discuss an ensemble feature set optimized for the target flipped classroom course. In the second part, we extend our analysis to courses with different educational settings (i.e. MOOCs), domains, and structure. Our results show that (i) the ensemble of optimal features varies depending on the course setting and structure, and (ii) the predictive performance of the optimal ensemble feature set highly depends on the course activities.