Temporal Convolutional Knowledge Tracking Model with Embedded Graph Association Information
Zhentao Zhong, Zhaohui Liu, Weifeng Gu · 2023
In recent years, deep knowledge tracking (DKT) and other deep learning-based knowledge tracking work has achieved relatively good results in terms of performance compared with traditional methods, but there are also some shortcomings: first, the existing methods fail to effectively use the deep information between question and knowledge points for effective mining, and the model interpretability needs to be improved; second, the original DKT model has long-term dependency problems, and it cannot use more previous interaction sequence to model students’ knowledge state more accurately. Based on this, a temporal convolutional knowledge tracking method with embeded graph association information is proposed in this paper, using graph attention network to mine the high-dimensional graph correlation information between knowledge points embedded in the input of the original model to enhance its interpretability, and then using temporal convolutional network (TCN) to extract students’ dynamically changing knowledge states, using expanded convolution and deep neural network to expand the sequence learning range and alleviate the long sequence dependence problem. The results show that the model outperforms traditional deep knowledge tracking methods in terms of AUC evaluation metrics under public dataset conditions, captures students’ knowledge states more accurately, and predicts students’ future performance more efficiently.