Mage-Kt Advances Knowledge Tracing by Modelling Inter-Concept Relations and Scale

Researchers are tackling the complex problem of accurately modelling a student’s learning path and predicting future performance, a field known as Knowledge Tracing. Chi Yu, Hongyu Yuan (Department of Computer Science, Inner Mongolia University) and Zhiyi Duan (Department of Computer Science, Inner Mongolia University) et al. present a significant advance by introducing MAGE-KT, a novel framework designed to better represent the relationships between students, questions and underlying knowledge concepts. Existing graph-based approaches often struggle with capturing nuanced inter-concept links and become overwhelmed by the sheer size of knowledge tracing graphs, leading to inaccurate predictions. MAGE-KT overcomes these limitations by constructing a multi-view graph, retrieving relevant subgraphs, and employing an innovative fusion module , ultimately delivering substantial improvements in both knowledge concept relation accuracy and next-question prediction across multiple datasets.

Existing graph-based approaches often struggle with capturing nuanced inter-concept links and become overwhelmed by the sheer size of knowledge tracing graphs, leading to inaccurate predictions.

MAGE-KT enhances knowledge tracing with multi-agent graphs

This innovative approach moves beyond simply inferring relationships from interaction sequences, instead leveraging a collaborative multi-agent system to generate, score, and adjudicate five distinct types of inter-knowledge concept relations, predecessor-successor, sibling, equivalence, containment, and association, ensuring both validity and structural consistency. This subgraph retrieval process focuses computational resources on the most relevant information, avoiding the attention diffusion and irrelevant computation that plague many existing methods. This breakthrough reveals a powerful new approach to knowledge tracing that not only enhances predictive accuracy but also addresses critical computational challenges. The integration of a multi-agent system for knowledge concept relation extraction ensures a more robust and semantically sound representation of the underlying knowledge structure.
By selectively retrieving relevant subgraphs and employing an asymmetric fusion module, MAGE-KT avoids the pitfalls of full-graph encoding, reducing noise and improving efficiency. The research establishes a strong foundation for future advancements in personalised learning systems, offering the potential to create more effective and adaptive educational experiences for students of all levels. The study unveils a framework that couples relation quality with efficient computation through three mutually reinforcing designs, beginning with the construction of a heterogeneous graph. Subsequently, a student-conditioned subgraph retriever leverages the target student’s history to jointly select high-value subgraphs from both views, concentrating computation where it matters and avoiding irrelevant attention diffusion.

Multi-agent knowledge graph construction for tracing learning

This innovative approach aims to enhance prediction accuracy while minimising computational cost and noise. The study pioneered a heterogeneous graph construction pipeline, integrating a multi-agent system for extracting knowledge concept relationships. Five distinct types of inter-concept relations were generated, scored, and adjudicated by specialised agents, ensuring both semantic validity and structural consistency within the knowledge graph. Simultaneously, a student-question interaction graph was created, incorporating Item Response Theory (IRT)-derived abilities and difficulties to model personalised student-question dynamics.

This dual-graph approach captures complementary dimensions of knowledge and learning behaviour, providing a richer representation of the learning process. Experiments employed a student-conditioned subgraph retriever to focus computation on the most relevant information. Given a target student’s history, the retriever jointly selects high-value subgraphs from both the knowledge concept graph and the student-question interaction graph. This targeted selection avoids attention diffusion into student-irrelevant regions, significantly reducing computational burden and improving the fidelity of inter-knowledge concept relations.

Results showed clear gains in next-question prediction compared to existing methods, confirming the effectiveness of the proposed framework. The innovative combination of multi-agent relation extraction, subgraph retrieval, and asymmetric fusion enables more accurate and efficient knowledge tracing, paving the way for more personalised and effective learning experiences. The approach achieves a significant step forward in modelling complex student learning trajectories.

MAGE-KT improves knowledge tracing with graph networks

The team measured performance across three widely used KT datasets: ASSIST09, Junyi, and Statics2011, revealing significant gains over existing methods. On the ASSIST09 dataset, MAGE-KT attained an accuracy (ACC) of 83.06 and an area under the curve (AUC) of 87.89, representing the highest values achieved amongst the compared models. Furthermore, on the Junyi dataset, the model recorded an ACC of 90.33 and an AUC of 91.79, again surpassing all baselines, these results indicate robust generalisation across different datasets. The breakthrough delivers a two-stage, direction-sensitive design that strengthens KC grounding for each instance while minimising unnecessary attention spread.

Results demonstrate that MAGE-KT’s directed information flow suppresses redundant attention and ineffective interactions, yielding more stable performance with reduced overfitting. Specifically, the researchers implemented a GRU with a hidden size of 512 and a dropout rate of 0.3 for sequence modelling, and stacked three layers of asymmetric cross-attention modules, each with four heads and a model dimension of 128, for multi-view fusion. Training employed the Adam optimiser with a learning rate of 1e-3 and a weight decay of 1e-5, utilising a batch size of 64 and early stopping on validation AUC with a patience of 10, for up to 100 epochs. Tests prove that removing any component of MAGE-KT consistently lowers ACC/AUC across all datasets, confirming the necessity and complementarity of its designs. Ablation studies on the Junyi dataset, utilising curated prerequisite, successor relations, showed the full pipeline achieved 92.52% prediction, 91.73% correctness, and a Jaccard index of 85.46, significantly outperforming the single-agent variant and ablations without completion or correction. Measurements confirm that the multi-agent relationship extraction pipeline effectively enriches KC descriptions and resolves spurious relationships, enhancing the model’s ability to accurately trace student learning trajectories.

MAGE-KT captures semantic and behavioural learning signals

The authors acknowledge that their method, while achieving state-of-the-art results, could be further extended to continual learning settings to adapt to real-time data. Future research will also explore incorporating richer textual information to refine knowledge concept relations and investigate uncertainty calibration for improved reliability in practical applications. These advancements promise to refine the modelling of student knowledge and enhance the effectiveness of educational tools, though the current work represents a substantial step forward in the field of knowledge tracing.