MOOC dropout prediction using a fusion deep model based on behaviour features

Abstract

High dropout rates have always been a problem for massive open online courses (MOOCs) to address. Dropout prediction has been used as an important method to address this situation. Learners' online learning behaviour is a good indicator of MOOC learning status. However, traditional prediction models treat each learning behaviour as an independent variable and ignore the correlation between learning behaviours. To address this issue, we propose a fusion deep dropout prediction model. The convolutional neural network (CNN) is used to obtain local feature representations based on learners' behaviour data, and the self-attention mechanism is used to learn correlations between different feature representations. At the same time, a bidirectional long short-term memory network (Bi-LSTM) is used to obtain a time-series feature vector representation and input the vector representation into the sigmoid function for prediction. We use the fusion deep model to predict behaviour data of 55,045 learners on the KDD Cup 2015 dataset. The experimental results show that the accuracy of the fusion deep model is 3% higher than that of the traditional method, which effectively improves the accuracy of dropout prediction.