Leveraging Knowledge Graphs and Large Language Models to Track and Analyze Learning Trajectories

Abstract

Learning is a dynamic process, yet traditional assessment captures only point-in-time snapshots. We combine knowledge graphs with LLMs to automatically build learning trajectories from student–AI dialogue, making the construction of student knowledge visible and tractable across an entire semester.

Problem & Motivation

Learning is a dynamic process — students' understanding of concepts shifts continuously over a semester — yet traditional assessment captures only point-in-time snapshots. Instructors have lacked tools to trace student knowledge construction across the full course timeline.

Method

We propose a method that combines knowledge graphs with large language models to automatically construct learning trajectories from student–AI tutor dialogue. Knowledge graphs represent inter-concept relationships; the LLM extracts knowledge structure from natural-language exchanges.

Findings

• Knowledge graphs are effective at visualizing learning trajectories and concept-mastery progression.
• LLMs can extract knowledge structure from dialogue logs with sufficient accuracy.
• Trajectory analysis enables early identification of struggling students.
• Accepted to LAK25 (Learning Analytics & Knowledge), a leading venue in the field.

Implications

Visualizing learning trajectories lets instructors see how each student is constructing knowledge — which concepts are stuck, which connections have yet to form. This makes targeted personalized tutoring practically actionable rather than aspirational.

Citation

@inproceedings{chen2025knowledge_graph,
  author    = {Yu-Hsuan Chen and Jhih-Sheng Huang and Jia-Yi Hung and Chia-Kai Chang},
  title     = {Leveraging Knowledge Graphs and Large Language Models to Track and Analyze Learning Trajectories},
  booktitle = {Int. Conf. on Learning Analytics \& Knowledge (LAK)},
  year      = {2025},
  month     = mar,
  doi       = {10.48550/arXiv.2504.11481},
  note      = {Also available at arXiv:2504.11481},
}