Answers are grounded in this paper and cite their sources.
Knowledge tracing aims to model students’ evolving knowledge states and predict future learning performance from historical interactions. Recent collaborative knowledge tracing methods complement individual histories with information from other students, but peer retrieval and evidence aggregation are often based on behavioral or sequence-level similarity without explicitly considering whether the corresponding interactions occur under comparable difficulty conditions. To address this limitation, we propose Difficulty-aware Collaborative Knowledge Tracing (DaCKT), a unified framework that uses item difficulty as a conditioning signal for cross-student collaborative modeling. Specifically, DaCKT constructs difficulty-conditioned behavioral profiles for coarse-grained peer retrieval, extracts collaborative evidence under same-item or difficulty-comparable conditions, and employs difficulty-aware attention to regulate the aggregation of individual and peer information. In addition, a difficulty-constrained supervised contrastive objective regularizes representations according to response relationships within locally comparable difficulty neighborhoods. Experiments on four public datasets show that DaCKT achieves the best AUC and F1 results among the compared methods across all datasets, while obtaining the best ACC on the three ASSISTments datasets. Controlled experiments under limited target history further show that DaCKT maintains higher AUC than representative sequential, difficulty-aware, and collaborative baselines across the examined history lengths, while a separate training-data sparsity analysis demonstrates gradual performance degradation as the available training information is reduced.
No takes yet. Share an insight, caveat, or question.
Lü et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: