Abstract
<jats:p>To effectively determine students' levels of mastery and to make customized educational suggestions, adaptive learning systems rely on intelligent knowledge tracing mechanisms. Traditional sequential-based and graph-based techniques for modeling learning are limited in their ability to model long-range conceptual dependencies, diverse education relationships, and sparse student interactions. This paper proposes a Graph Transformer–Based Self-Supervised Learning Framework for modeling students’ knowledge tracing in adaptive learning environments that provides a novel approach to addressing the above limitations. Through the creation of heterogeneous educational graphs from students' interactions with exercises, the new framework creates graph transformers to learn the contextual relationships between students, concepts, exercises, and the temporal progression of their learning. Furthermore, a self-supervised contrastive learning approach to learning latent representations improves both robust and generalizable learning due to incomplete education records. The framework developed was evaluated using benchmark datasets for adaptive learning containing records of interaction, response correctness, concept dependencies, and temporal learning behaviours. This paper presents a framework for knowledge representation learning and predictive analytics that combines self-supervised contrastive optimization with graph transformer representation learning. The experimental results indicate that the framework achieves a Knowledge State Estimation Rate of 96.8%, a Learning Path Adaptation Score of 94.7%, an Interaction Stability Index of 93.1%, a Graph Representation Consistency of 92.6%, and a Recommendation Optimization Factor of 91.4%, outperforming the current state-of-the-art methods including Deep Knowledge Tracing, Dynamic Key-Value Memory Networks, Graph Neural Knowledge Tracing, and Transformer Knowledge Tracing. Also, the proposed framework has demonstrated better capabilities in terms of contextual representation learning and adaptive recommendation through graph attention and self-supervised optimization. Thus, this research confirms that the proposed framework can serve as an efficient and scalable approach to intelligent educational systems by improving student performance predictions, adaptive recommendation quality, and learning path optimization.</jats:p>