Deep Knowledge Tracing Model with an Evolved Transformer Structure
Zhijun Li, Zixiao Xue, Chen Liu, Yanzhang Feng · 2023
Deep Learning based Knowledge Tracing (DKT) has become a research hotspot in the intelligence education field. Compared to conventional methods, DKT has better predictive performance, but it also has some problems such as poor interpretability, and difficulties in reflecting the causal association between the learning process and test results. In this paper, a new DKT model is proposed based on an evolved Transformer structure (DKT-ETS). The encoder layer is composed of three coding networks with a multi-head self-attention mechanism, while inputs are three types of pre-processed data: process characteristic data, test label data, and answer results data. Output as three matrices of V, Q, K. The decoder layer also uses the attention mechanism, in which the input is the three matrices that come from encoder, and the output is the predicted result. By improving the structure, the new model introduces certain interpretability into the V, Q and K matrices of the attention mechanism. Thus, the causal relationship between the learning process and test results can be reflected a certain extent: the V matrix represents the characteristic information of the testee's learning process; the Q matrix reflects the knowledge point information examined by the current test item; and the K matrix represents the results of the previous tests. DKT-ETS was validated by using the large-scale knowledge tracking data set EdNet, and the results show that its ACC and AUC evaluation indicators have been significantly improved.