Contrastive Deep Knowledge Tracing

Huan Dai, Yue Yun, Yupei Zhang, Wenxin Zhang, Xuequn Shang

Abstract: Knowledge tracing (KT) aims to predict student performance on the next question according to historical records. Recently deep learning-based models for KT task successfully modeling student responses receive good prediction results of student performance. The student responses encoded as input of KT models use a one-hot encoding. We find that one-hot encoding represents student responses on different items related to the same concepts in completely different vectors. However, items related to the same concept have certain relationships in the real world so the student has a similar representation in these items. In this paper, we propose a new method named Contrastive Deep Knowledge Tracing (CDKT) for providing a reasonable representation of students. We evaluate our model using three public benchmark datasets and the experimental results demonstrate improvements over state-of-the-art methods.