Data Augmentation for Knowledge Tracing Based on Variational AutoEncoder and Efficient Network Reusing

Knowledge tracking (KT) is a task that predicting the degree of students’ knowledge mastery through their learning interaction records. Although existing works improve predictive capability with well-designed neural network models or hypothetical learning mechanisms, the predictive performance is compromised in the scenarios of quantity limited interaction data. In this paper, we utilize Variational AutoEncoder (VAE) and pre-trained network to generate question-answer sequence data pairs related to the original interaction data, which can improve the performance of the model when added to the training set even in the case of data scarcity. Specifically, the steps of the data augmentation method for KT we proposed are as follows: 1) Question sequence generation. Generate latent question sequences that are similar to the real interaction question sequences from the pre-designed VAE model. 2) Answer sequence generation. Put the generated data into the pre-trained KT model to get reliable answer label sequences that correspond to latent question sequences. 3) Samples generation and training. Combine the two types of generated sequences as new samples for KT task training. We apply the data augmentation method on four classic datasets and demonstrate its effectiveness by reaching the state-of-the-art performance with an average AUC index improvement of 2.41%. We also verify the method on artificially random extracted data, and with only 20% of the data, it even achieves similar results compared with other methods using 100% of the data.