FuMoE-csKT: fusing disentangled and self-attention in a personalized mixture-of-experts framework for cold-start knowledge tracing

Knowledge tracing (KT) aims to model a student’s evolving mastery of knowledge components (KCs) from historical interaction records. However, real-world online learning systems inevitably face cold-start settings, where new students enter the platform with only limited interaction data. Although numerous deep learning based KT models have been proposed, achieving robust performance under cold-start conditions remains challenging. In this article, we propose FuMoE-csKT, a personalized mixture-of-experts (MoE) framework for cold-start KT that integrates disentangled attention with self-attention. At the core of FuMoE-csKT is a fused attention module that jointly models disentangled attention and self-attention in a unified architecture and is further enhanced by a learnable kernel-bias function. This design strengthens sequence modeling and temporal robustness, particularly for capturing multi-dimensional relations in short interaction sequences. To further improve representation diversity and personalization, we feed the attention outputs into a soft MoE layer, where a gating mechanism softly selects multiple expert networks. Such expert diversity enables the model to adapt to heterogeneous student behavior patterns even under severe data sparsity. We conduct comprehensive evaluations on four real-world educational datasets under cold-start conditions using two widely adopted metrics, area under the curve (AUC) and accuracy (ACC). The results show that the proposed framework consistently outperforms existing baselines in most cases. We additionally perform ablation studies and efficiency analyses to validate the effectiveness and robustness of our framework.