Adaptive Knowledge Transfer for Cross-Disciplinary Cold-Start Knowledge Tracing

TL;DR

This paper tackles cross-disciplinary cold-start knowledge tracing (CDCKT), where target-discipline data are severely limited. It introduces ACKT, a framework that combines a category-guided mixture-of-experts mapping (CMOE) with adversarial preference distribution alignment to transfer knowledge from a data-rich source discipline to a sparse target discipline. The method pre-trains on the source data, clusters source knowledge states into K categories, and uses category information to gate multiple experts, while an adversarial discriminator leverages non-overlapping data to refine cross-domain mappings. Across 20 extreme cross-disciplinary scenarios on five real-world datasets, ACKT consistently outperforms baselines, demonstrates robustness under very low overlap, and shows favorable scalability for deployment. This work significantly advances CDCKT by enabling effective cross-domain transfer without heavy reliance on overlapping entities, with practical implications for personalized education in new subjects.

Abstract

Cross-Disciplinary Cold-start Knowledge Tracing (CDCKT) faces a critical challenge: insufficient student interaction data in the target discipline prevents effective knowledge state modeling and performance prediction. Existing cross-disciplinary methods rely on overlapping entities between disciplines for knowledge transfer through simple mapping functions, but suffer from two key limitations: (1) overlapping entities are scarce in real-world scenarios, and (2) simple mappings inadequately capture cross-disciplinary knowledge complexity. To overcome these challenges, we propose Mixed of Experts and Adversarial Generative Network-based Cross-disciplinary Cold-start Knowledge Tracing Framework. Our approach consists of three key components: First, we pre-train a source discipline model and cluster student knowledge states into K categories. Second, these cluster attributes guide a mixture-of-experts network through a gating mechanism, serving as a cross-domain mapping bridge. Third, an adversarial discriminator enforces feature separation by pulling same-attribute student features closer while pushing different-attribute features apart, effectively mitigating small-sample limitations. We validate our method's effectiveness across 20 extreme cross-disciplinary cold-start scenarios.

Figures (5)

• Figure 1: The overview Architecture of ACKT. It consists of two stages, where stage 1 is to extract high-quality knowledge states and question embeddings, and to cluster the knowledge states, as shown in (a), (b). Stage 2 mainly consists of a category-enhanced mixture of experts mapping module (CMOE) and unsupervised preference distribution adversarial optimization, as shown in (c), (d).
• Figure 2: Feature extraction process in the prediction phase for general knowledge tracing methods
• Figure 3: Effect of model parameters on performance
• Figure 4: The effect of the number of overlapping students on model performance.
• Figure 5: Visualization of students' knowledge states under different circumstances.