Scholarlink

<p>The primary objective of knowledge tracing is to dynamically predict learners’ mastery of future exercises by analyzing their historical learning records, thereby facilitating the design of personalized learning paths. Traditional knowledge tracing models exhibit limitations in capturing long-term dependencies, handling noisy data, and modeling complex learning behaviors, which hinder their ability to effectively extract global features and maintain robustness. This study proposes a time-convolution knowledge tracing method based on a discrete cosine transform attention mechanism (DA-TCKT). The model captures long-term dependencies through a time-convolution network and enhances feature extraction by incorporating a frequency-enhanced channel attention mechanism grounded in the discrete cosine transform. Furthermore, the model integrates scalar long short-term memory networks and residual networks. The model achieves a maximum AUC improvement of 30.87% over the weakest baseline, demonstrating its effectiveness in adaptive knowledge tracing tasks.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

DA-TCKT: Adaptive time convolution knowledge tracing method based on discrete cosine transform attention mechanism

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年10月15日

查看全文
  • Hairui Wang,
  • Yunting Wang,
  • Guifu Zhu

摘要

The primary objective of knowledge tracing is to dynamically predict learners’ mastery of future exercises by analyzing their historical learning records, thereby facilitating the design of personalized learning paths. Traditional knowledge tracing models exhibit limitations in capturing long-term dependencies, handling noisy data, and modeling complex learning behaviors, which hinder their ability to effectively extract global features and maintain robustness. This study proposes a time-convolution knowledge tracing method based on a discrete cosine transform attention mechanism (DA-TCKT). The model captures long-term dependencies through a time-convolution network and enhances feature extraction by incorporating a frequency-enhanced channel attention mechanism grounded in the discrete cosine transform. Furthermore, the model integrates scalar long short-term memory networks and residual networks. The model achieves a maximum AUC improvement of 30.87% over the weakest baseline, demonstrating its effectiveness in adaptive knowledge tracing tasks.