A Learning Style-aware Dropout Prediction Method via Fusing Global and Local Semantics

Massive Open Online Courses (MOOCs), with their vast resources and significant flexibility, have emerged as the most popular way for online education globally. However, one of the emergent challenges faced by numerous MOOC platforms is the high rate of dropout. While many approaches have been proposed to predict dropout rates, the common drawback is that they usually ignore the theories from educational psychology when modeling the behavior from log data, deriving undesirable performance for the task of dropout prediction. In this paper, we propose a novel learning style-aware framework for predicting dropout on MOOCs, effectively fusing global and local semantics contained in learners’ learning behaviors. First, the learning styles are identified by clustering on time series data in logs of MOOCs. Then, the global and local semantics contained in the learning style information and learning behavior information are captured by a module of CNNs, deriving the meaningful hidden representations. Finally, the obtained high-quality representations are used to derive the predicted dropout behavior given learner-course pairs. Extensive experiments are carried out on the KDD Cup 2015 dataset, and the results demonstrate the effectiveness of the proposed method.