Attentive pre-training question embeddings for knowledge tracing with semantically-enhanced knowledge structure and concept label-guided heterogeneous graph representation.

Abstract

Personalized intelligent education's popularity has fueled demand for accurate prediction of learner performance through Knowledge Tracing (KT). However, current concept-level KT methods mainly treat all concepts associated with questions of varying difficulty levels equally, resulting in a lack of specificity in capturing question-related information. Additionally, some approaches learn question embeddings during model training, which can lead to a complex entanglement of question embeddings with the model. To address these issues, our study proposes an attentive pre-training embedding method called Semantic information Retrieval augmentation and Concept Label-Heterogeneous Graph representation for KT (SRHGKT). This method directly learns the interaction between questions and concepts through specialized designs, such as devising hybrid semantic retrieval to construct a knowledge structure that captures rich information about concepts. Furthermore, we design an innovative concept label-guided heterogeneous graph embedding fusion module to combine the advanced information from question-concept interactions with multiple aspects, resulting in pre-training question embeddings. We also introduce a forgetting question similarity attention to model the forgetting patterns of learners. Comprehensive experimental results, conducted on three real-world datasets, demonstrate that SRHGKT outperforms 14 state-of-the-art methods in predicting learner performance. Furthermore, generalizability tests show that our pre-training embeddings exhibits impressive generalization, achieving improvements of over 10% prediction accuracy, when applied to various KT models on the ASSISTment2009 dataset.