Sequence Denoising with Self-Augmentation for Knowledge Tracing

Shanshan Wang, YING HU, Xun Yang, Ke Xu, Mengzhu Wang, Yuanhong Zhong, Xingyi Zhang

Keywords: knowledge tracing，sequence denoising，data augmentation，ai for education

Abstract: Knowledge tracing (KT) aims to predict students' future knowledge levels based on their historical interaction sequences. Most KT methods rely on interaction data between students and questions to assess knowledge states and these approaches typically assume that the interaction data is reliable. In fact, on the one hand, factors such as guessing or slipping could inevitably bring in noise in sequences. On the other hand, students' interaction sequences are often sparse, which could amplify the impact of noise, further affecting the accurate assessment of knowledge states. Although data augmentation which is always adopted in KT could alleviate data sparsity, it also brings noise again during the process. Therefore, denoising strategy is urgent and it should be employed not only on the original sequences but also on the augmented sequences. To achieve this goal, we adopt a plug and play denoising framework in our method. The denoising technique is adopted not only on the original and the enhanced sequences separately during the data augmentation process, but also we explore the hard noise through the comparison between the two streams. During the denoising process, we employ a novel strategy for selecting data samples to balance the hard and soft noise leveraging Singular Value Decomposition (SVD). This approach optimizes the ratio of explicit to implicit denoising and combines them to improve feature representation. Extensive experiments on four real-world datasets demonstrate that our method not only enhances accuracy but also maintains model interpretability.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 919