Embedding cognitive framework with self-attention for interpretable knowledge tracing.

Area of Science:

• Educational Data Mining
• Artificial Intelligence in Education
• Cognitive Modeling

Background:

• Deep neural networks, like deep knowledge tracing, are used in learning analytics but lack interpretability.
• Understanding latent student learning status is crucial for effective educational interventions.

Purpose of the Study:

• To propose the Enhanced Cognitive Knowledge Tracing (EAKT) framework for improved learner modeling.
• To enhance the interpretability of cognitive models in educational data mining.
• To provide a multi-dimensional view of student knowledge states.

Main Methods:

• Embedding a structured cognitive model within a transformer architecture.
• Developing the EAKT framework to analyze fine-grained student learning processes.
• Conducting performance and interpretability experiments on real-world datasets.

Main Results:

• EAKT achieved superior prediction of student learning outcomes, exceeding baseline methods by over 2.6% on two datasets.
• Interpretability experiments showed EAKT's student knowledge state tracing is closer to ground truth.
• The framework demonstrated enhanced accuracy in tracking changes in student knowledge states.

Conclusions:

• EAKT offers a powerful and interpretable learner modeling framework.
• The model enhances the design and implementation of intelligent tutoring systems.
• EAKT advances the field of educational data mining with its dual predictive and explanatory capabilities.