SKG-KT: Semantic Knowledge Graph Construction and Reasoning for Knowledge Tracing by Large Language Models
TL;DR
A novel KT framework that leverages the reasoning ability of large language models (LLMs) to enhance the modelling of unseen students, without relying on numerous students with historical records is proposed and validated to achieve robust predictive performance compared to others.
Abstract
Knowledge Tracing (KT), a core component of intelligent tutoring systems, aims to model the dynamic evolution of students' knowledge states based on their historical interaction data. However, mainstream methods rely heavily on large-scale students' response logs, hindering generalization to unseen students, and often overlook the semantic structure of knowledge points in knowledge graph construction. In this paper, we propose Semantic Knowledge Graph-Based Knowledge Tracing (SKG-KT), a novel KT framework that leverages the reasoning ability of large language models (LLMs) to enhance the modelling of unseen students, without relying on numerous students with historical records. SKG-KT consists of two stages: ATC-ruled Semantic Graph Extraction and a Concept-aware Reasoning mechanism. In the first stage, we design a semantic student knowledge graph extraction method grounded in ATC rules, which captures the semantic structure across three hierarchical layers—domain, cluster, and standard—aligned with established educational theory. To extend beyond mathematics, we further introduce a multi-disciplinary semantic graph extraction method that induces an ATC-inspired hierarchical schema for other disciplines via LLM-based label induction. In the second stage, the Concept-aware Reasoning mechanism first applies knowledge concept reasoning to infer key knowledge concepts for unseen students with limited previous student data, and then fine-tunes an open-source LLM for correctness prediction based on the reasoned knowledge concepts, adapting the LLM to the KT task. Extensive experiments conducted on three publicly available educational benchmark datasets, namely Xes3g5m-small, Xes3g5m-large, and MOOCRadar, demonstrate that SKG-KT achieves state-of-the-art performance in knowledge tracing. Furthermore, SKG-KT has been validated to achieve robust predictive performance compared to others, while relying on a limited number of students with historical records. Our code and prompts are publicly available at the link . https://github.com/linqingliDestinyController/SKG_KT/tree/main.
