Scholarlink

<p>Accurately modeling learners’ knowledge states is crucial for advancing personalized intelligent education. However, existing knowledge tracing methods often overlook the influence of problem complexity on students’ answering strategies, leading to unstable and inaccurate predictions. To address these challenges, we propose PSKT, a novel deep knowledge tracing model that integrates problem complexity and state stability. PSKT incorporates (1) a quantitative representation of problem complexity using information entropy and accuracy, (2) dynamic adjustments of knowledge states based on perceived and actual problem difficulty, and (3) a contrastive learning-based mechanism to stabilize predictions and reduce information bias. Experiments on four public datasets–ASSIST2009, ASSIST2015, Algebra05, and Statics2011–demonstrate that PSKT outperforms six state-of-the-art models, achieving up to 3.67% higher AUC and improved robustness across all datasets. These results highlight the potential of PSKT to enhance predictive performance and provide more reliable insights into students’ learning processes, making it a valuable tool for personalized education systems.</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

A Novel Deep Knowledge Tracing Model with Problem Complexity and State Stability

  • 期刊论文

  • 发表日期：

    2025年7月15日

查看全文
  • Xinxin Li,
  • Fei Luo,
  • Junhai Ouyang,
  • Luis Rojas Pino,
  • Wenhai Li,
  • Weichao Ding,
  • Chunhua Gu

摘要

Accurately modeling learners’ knowledge states is crucial for advancing personalized intelligent education. However, existing knowledge tracing methods often overlook the influence of problem complexity on students’ answering strategies, leading to unstable and inaccurate predictions. To address these challenges, we propose PSKT, a novel deep knowledge tracing model that integrates problem complexity and state stability. PSKT incorporates (1) a quantitative representation of problem complexity using information entropy and accuracy, (2) dynamic adjustments of knowledge states based on perceived and actual problem difficulty, and (3) a contrastive learning-based mechanism to stabilize predictions and reduce information bias. Experiments on four public datasets–ASSIST2009, ASSIST2015, Algebra05, and Statics2011–demonstrate that PSKT outperforms six state-of-the-art models, achieving up to 3.67% higher AUC and improved robustness across all datasets. These results highlight the potential of PSKT to enhance predictive performance and provide more reliable insights into students’ learning processes, making it a valuable tool for personalized education systems.