Attentive pre-training question embeddings for knowledge tracing with semantically-enhanced knowledge structure and...

Area of Science:

• Educational Technology
• Artificial Intelligence in Education
• Machine Learning for Learning Analytics

Background:

• Personalized intelligent education necessitates accurate learner performance prediction via Knowledge Tracing (KT).
• Existing concept-level KT methods often lack specificity, treating concepts equally regardless of question difficulty.
• Some KT approaches entangle question embeddings within the model during training, complicating the learning process.

Purpose of the Study:

• To develop an advanced pre-training embedding method for Knowledge Tracing (KT) that addresses limitations in current concept-level approaches.
• To enhance the specificity and accuracy of learner performance prediction in intelligent education systems.
• To create generalizable pre-trained embeddings for KT that can improve various existing models.

Main Methods:

• Proposed Semantic information Retrieval augmentation and Concept Label-Heterogeneous Graph representation for KT (SRHGKT).
• Employed hybrid semantic retrieval to build a knowledge structure capturing rich concept information.
• Designed a concept label-guided heterogeneous graph embedding fusion module for pre-training question embeddings.
• Introduced forgetting question similarity attention to model learner forgetting patterns.

Main Results:

• SRHGKT outperformed 14 state-of-the-art methods in learner performance prediction across three real-world datasets.
• Pre-trained embeddings demonstrated significant generalizability, improving prediction accuracy by over 10% on the ASSISTment2009 dataset when applied to various KT models.
• The method effectively captures question-concept interactions and learner forgetting patterns.

Conclusions:

• SRHGKT offers a more specific and accurate approach to Knowledge Tracing compared to existing methods.
• The proposed pre-training embeddings enhance the performance and generalizability of diverse KT models.
• This research contributes to more effective personalized intelligent education systems through improved learner modeling.