Knowledge Tracing (KT) aims to assess students' dynamic knowledge states from their learning histories. Most existing approaches focus on single‑domain learning, yet real‑world education often involves multiple domains simultaneously, introducing two critical factors: cognitive load and knowledge transfer. Cognitive load refers to the mental effort of managing learning across domains in both temporal and knowledge dimensions, while knowledge transfer describes how a knowledge state in one domain influences related states within and across domains.
This paper presents LT-MKT (Load and Transfer for Multi-domain Knowledge Tracing), a novel method designed for multi‑domain scenarios. First, textual information from questions and their associated concepts is encoded using large language models (LLMs) to build a multi‑domain hierarchical graph, bridging otherwise isolated domains. Then, cross‑domain features are explicitly modeled along temporal and knowledge dimensions to capture the effects of cognitive load. A knowledge transfer module is further introduced to propagate knowledge states both within and across domains. By jointly learning these components, LT‑MKT achieves more accurate predictions of students' future performance.
Extensive experiments on real‑world datasets demonstrate that LT‑MKT consistently outperforms existing baselines, achieving state‑of‑the‑art results.
Blogger's Review: The study ingeniously integrates LLM‑driven semantic representations with a cross‑domain graph structure, systematically addressing cognitive load and knowledge transfer in multi‑domain KT. The strong empirical results suggest great potential for deployment in intelligent education systems.