A CNN-Transformer Knowledge Tracking Model Combining with Forgotten Factors
Xiaosheng Huang, Yin Wu, Yixin Luo · 2024
Knowledge tracking plays a crucial role in intelligent education, contributing significantly to promoting educational equity and enhancing educational quality. To address the limitations of the Transformer-based knowledge tracking model in terms of forgetting characteristics and local modeling, this paper proposes a knowledge tracking approach that integrates convolutional neural network (CNN) and Transformer, drawing on the Ebbinghaus forgetting curve theory and memory trace decay theory. This paper incorporates three forgetting factors into the Transformer model: the time interval of learning, the number of times a learner revisits the same knowledge points, and the learner's mastery of knowledge points to address its deficiencies in forgetting characteristics. Simultaneously, one-dimensional convolution is used to model the learner's historical interaction sequence to address the relative shortcomings of the Transformer model in local modeling. Finally, the outputs of the global modeling Transformer layer and the local modeling convolution layer are concatenated to obtain the learner's knowledge state level, thus enabling predictions of the learner's future performance. Experiments on four main datasets show that compared with BKT, DKT, DKVMN, SAKT and SAINT, the proposed model has achieved a significant improvement in AUC performance and running time.