Prediction of online learning behavior in universities based on long short-term memory networks

Abstract

If potential problems in students’ online learning can be revealed, they can optimize their learning methods through self-correction, thereby effectively improving learning outcomes. However, since students often find it difficult to discover bad habits in their studies on their own, it is necessary to use neural network tools to achieve this goal. Traditional prediction neural networks typically use time series analysis methods, which have the disadvantage of being difficult to effectively capture long-term dependencies in learning behavior. However, the double-layer long short-term memory (LSTM) network designed in this article can solve the problem of long-term dependencies. Users’ online learning datasets provided by online learning platforms Kaggle, edX, and the Open University of the UK are used to compose the dataset used in this paper, and pre-processing such as serialization, format unification, smoothing, and data filtering is performed before training. The TensorFlow framework is used in terms of writing programs in python language to build the double-layer LSTM network model. The batch size of the training dataset is set to 100, and a total of 50 rounds of training are performed. The training loss finally converges to 0.0712, and the testing loss ranges from 0.1 to 2.0. After training, this paper conducts two sets of experiments to evaluate the model performance. In the first set of experiments the average losses of the double-layer LSTM network, the single-layer LSTM network, the BPNN, and the RBFNN on the same testing dataset are compared, which are 1.1044, 1.5545, 1.8921, and 1.8494, respectively. The results show that the double-layer LSTM network has better performance in the models. In the second set of experiments, the double-layer LSTM network is applied to predicting grades on the respective user datasets of Coursera, Udacity, and edX, and its accuracy is compared with the platforms’ own prediction accuracy. Among them, Coursera has an accuracy rate of 85.3% on its own user data; the accuracy of double-layer LSTM on Coursera user data is 83.4%; Udacity’s accuracy on its own user data is 80.1%; the accuracy of double-layer LSTM on Udacity user data is 76.6%; the accuracy of edX on its own user data is 81.9%; the accuracy of double-layer LSTM on edX user data is 81.4%. The results show that none of the predictions of the double-layer LSTM networks are as accurate as the platforms, and the analysis indicates that it is probably related to the optimization of the platforms for their own users.