Article Preview
TopIntroduction
The widespread adoption of online education platforms has enabled continuous recording and quantitative analysis of students’ learning behaviors in digital environments (Jung et al., 2025; Kuo, 2025; Kuo et al., 2025). From a data warehousing and mining perspective, these platforms can be viewed as domain-specific educational interaction warehouses that continuously collect student IDs, item IDs, concept tags, response outcomes, timestamps, and behavioral traces. Therefore, the central analytical task is to mine predictive and stable temporal-relational patterns from large-scale heterogeneous records so that institutions can support early warning, resource allocation, personalized learning services, and learning analytics decision making. Thus, in this study, knowledge tracing is reformulated not only as a pedagogical prediction task but also as a sequential data mining problem over evolving educational interaction data (Fu et al., 2024; Guo et al., 2025; Zhao, 2024).
Existing knowledge tracing approaches mainly follow two paradigms: sequential modeling and graph-enhanced modeling. Sequential models encode historical interactions over time through recurrent architectures, self-attention mechanisms, or generative sequence models (Wang, 2025; X. Zhou et al., 2025). Although these methods are effective for response sequence modeling, structural dependencies among items and knowledge components are often implicitly compressed into model parameters rather than explicitly mined as evolving relations (H. Liu et al., 2024; Park & Kim, 2025). Graph-enhanced models characterize structural dependencies via knowledge graph propagation (Kuo et al., 2024; Scarlatos et al., 2025; Tato & Nkambou, 2025), but most of them rely on fixed adjacency matrices or globally shared relation weights. This creates a data-mining gap: Temporal patterns and structural relations are modeled through separate mechanisms, making it difficult to update relational interpretations when learning stages shift or long interaction sequences accumulate. When the model lacks explicit storage and constraints for historical structural trajectories, relational drift and interpretational inconsistency may emerge, thereby affecting long-term predictive stability and cross-scenario generalization ability (Ma et al., 2025; Mai et al., 2025).
Therefore, the main challenge addressed in this study is how to couple temporal dynamics and relational dynamics within a unified data-mining framework while maintaining structural semantic continuity. This challenge is closely related to sequential pattern mining, dynamic graph mining, and predictive analytics for decision-support systems, rather than to domain-specific pedagogical modeling alone (Jung et al., 2025; Xu et al., 2024).
To overcome the aforementioned limitations, this paper introduces designs at both the propagation mechanism and representation memory levels. First, a time-conditioned mechanism is incorporated into the structural propagation stage. Instead of relying on fixed adjacency weights, the graph propagation process is modulated by factors generated from the current temporal state, enabling structural relations to dynamically adapt to evolving learning states. This design embeds temporal information directly into the structural propagation operator, transforming structural dependencies into time-evolving relations and alleviating the representational fragmentation caused by temporally–structurally decoupled modeling.
Second, a relational memory mechanism is introduced along the temporal dimension to store relational prototypes formed in previous learning stages. Through differentiable retrieval and temporal decay, the memory is continuously updated and selectively accessed. This mechanism provides cross-stage structural context for the current temporal state, maintaining semantic continuity in long-sequence modeling and reducing the drift risk induced by recursive hidden-state updates.