Early Prediction of Students’ Performance Using a Deep Neural Network Based on Online Learning Activity Sequence

Xiao Wen, Hu Juan

Predicting students’ performance is one of the most important issues in educational data mining. In this study, a method for representing students’ partial sequence of learning activities is proposed, and an early prediction model of students’ performance is designed based on a deep neural network. This model uses a pre-trained autoencoder to extract latent features from the sequence in order to make predictions. The experimental results show that: (1) compared with demographic features and assessment scores, 20% and wholly online learning activity sequences can achieve a classifier accuracy of 0.5 and 0.84, respectively, which can be used for an early prediction of students’ performance; (2) the proposed autoencoder can extract latent features from the original sequence effectively, and the accuracy of the prediction can be improved more than 30% by using latent features; (3) after using distance-based oversampling on the imbalanced training datasets, the end-to-end prediction model achieves an accuracy of more than 80% and has a better performance for non-major academic grades.