Construction of Knowledge Tracing Model Based on Side Information

Abstract Intelligent Tutoring Systems are becoming more and more important in online education. Knowledge Tracing is a crucial part of ITS and aims to track students' knowledge status according to their historical performance and predict their future performance. Recent research shows that Deep Knowledge Tracing (DKT) models exhibit better performance than traditional models. However, the input of the DKT models depends on the ID and the total number of Students, Knowledge components, and Questions (SKQ), so the DKT models can not quickly be transferred to other datasets. Moreover, the application of the trained models is strictly limited when they are put into use due to data privacy. To solve these problems, this study proposes 14 features as side information (information that does not contain ID) related to SKQ and constructs a Knowledge Tracing model based on Side Information (SIKT). The input of this model neither depends on specific ID and the total number of SKQ nor involves students' private information to apply to all datasets. The results in continual learning with three real-world datasets as independent tasks show that: 1) The SIKT model performs better than the baseline models on three datasets and uses the least number of parameters. The side information of students and questions has a significant impact on the effectiveness of knowledge tracing. 2) In continual learning, the SIKT model has better performance on tasks with wider data distribution and can share knowledge across different tasks. 3)The SIKT model does not have the problem of catastrophic forgetting and learns the knowledge of both the original task and the new task at the same time when learning new tasks. 4) The performance achieved by the SIKT model in continual learning is comparable to that achieved by pooling all tasks at once and the model can generalize better as the number of training tasks increases.