Self-learning Tags and Hybrid Responses for Deep Knowledge Tracing

Abstract

Knowledge Tracing, as a classic task of evaluating student knowledge mastery and predicting performance by modeling student response sequences, has become one of the key motivations for stimulating the vigorous development of personalized online education. With the support of Recurrent Neural Network, Deep Knowledge Tracing (DKT) and its variants demonstrate remarkable KT performance on account of their excellent learning ability and knowledge state representation. However, the drawbacks of these models have gradually emerged with the surge in student interaction data scale and the variety of interaction forms, which include employing the pre-defined knowledge point tags and the single response feature input. In this paper, we advocate a brand-new LTDKT-HR model, which constructs exercise embedding mapped from exercise space to tag space by self-training and optimizes input by adding an effective feature of first response time to the original feature of whether answer is correct or not. Sufficient experiments on two open datasets prove that LTDKT-HR outperforms the general DKT in student performance prediction, that is, self-learning tags are superior to existing manual tags. In addition, the proposed model can dig out the influence between exercises, which provides the mathematical basis for further setting exercise relationship constraints.