Dynamic comprehensive difficulty knowledge cells based on KAN network and stable learning for knowledge tracing.

Area of Science:

• Artificial Intelligence
• Educational Technology
• Machine Learning

Background:

• Knowledge Tracing (KT) models predict student performance but often fail to account for dynamic changes in comprehensive difficulty.
• Existing KT methods struggle with accurately updating difficulty and tracing knowledge states due to factors like lucky guesses by students.

Purpose of the Study:

• To propose a novel approach, Dynamic comprehensive Difficulty Knowledge Cells (DDKC), that addresses the limitations of current Knowledge Tracing models.
• To improve the accuracy and interpretability of student knowledge state tracing by incorporating dynamic difficulty and mitigating spurious correlations.

Main Methods:

• Developed DDKC, featuring dynamic comprehensive difficulty knowledge cells utilizing a KAN network to adapt difficulty based on student mastery.
• Implemented a Stable Learning training strategy to reduce the impact of lucky guesses and focus on genuine knowledge state tracing.

Main Results:

• DDKC demonstrated superior performance compared to state-of-the-art models on multiple public educational datasets.
• The model effectively captured dynamic changes in comprehensive difficulty and successfully eliminated spurious correlations.

Conclusions:

• DDKC offers a significant advancement in Knowledge Tracing by dynamically adjusting for difficulty and handling student response noise.
• The proposed method provides interpretable insights into how comprehensive difficulty evolves and how spurious correlations are mitigated.