ONet – A Temporal Meta Embedding Network for MOOC Dropout Prediction

While massive open online courses (MOOCs) represent a promising approach to support education for all, their effectiveness is limited by high dropout rates. Based on the analysis that dropout is deeply correlated with the interaction patterns of a user, this paper proposes an attention meta embedding based deep temporal network to predict user dropout in MOOCs. The attention meta embedding gives an attention based weighted representation of different candidate embeddings and temporal modeling is done by Bi-GRU. An experimental evaluation of our solution on the KDDCUP 2015 and XuetangX datasets shows better results (AUC and F1 score) compared to both handcrafted feature techniques and all but one of the auto feature generated techniques. We obtain comparable results to the best previous method, but without having to use different architectures for different datasets.