Enhancing Question Representation for Knowledge Tracing via Adaptive Concept Relation Discovery and Structural–Semantic Co-Optimization

Knowledge tracing (KT) models the evolution of students’ knowledge states from historical learning interactions, serving as a fundamental mechanism for intelligent educational systems. The efficacy of KT models heavily relies on the quality of question representations. However, existing methods incorporating concept relations generally face two critical limitations: First, static relation graphs fail to capture dynamic cognitive trajectories, while shallow statistical approaches are highly vulnerable to sparse data and interaction noise. Second, the current models lack fine-grained semantic characterization, specifically failing to integrate concept difficulty to accurately evaluate the cognitive load of individual questions. To address these issues, we propose SSCKT, a novel framework that adaptively extracts latent concept relations and introduces concept associations to enhance question representation via structural–semantic co-optimization. Specifically, a concept relation discovery and aggregation module dynamically learns a directed acyclic graph (DAG) of concept relations from historical learning interactions, which is encoded via heterogeneous graph convolutions to yield structure-enhanced embeddings. Concurrently, a concept difficulty-guided semantic enhancement module utilizes multi-head attention to capture intrinsic question–concept dependencies. This module employs a topological concept difficulty metric as a scaling factor to adjust the question offset vector, enabling cognitive load-aware modeling. Finally, a contrastive learning mechanism aligns the structural and semantic representations to eliminate feature redundancy, generating highly discriminative question embeddings. Extensive experiments on five public datasets demonstrate that, relative to the strongest baseline on each dataset, SSCKT improves average ACC and AUC by 1.20% and 1.33%, respectively, validating the superiority of the proposed co-optimization architecture.