CourseGraph: Finding overlaps and differences in Computer Science courses across universities

Student mobility programs like Erasmus+ offer students valuable international experience, but they create a significant administrative burden: ensuring that courses taken abroad do not substantially overlap with the student's home curriculum. Currently, this evaluation is performed manually by academic advisors, a process that is time-consuming and difficult to scale. The authors propose CourseGraph, an automated methodology designed to assist in this evaluation by comparing the educational content of external courses against a home university's curriculum.

The approach follows a four-step pipeline: (1) data extraction from course webpages using a Large Language Model (LLM) to standardize information like titles, descriptions, and learning outcomes; (2) semantic embedding of these components using sBERT; (3) computing pair-wise cosine similarity between course components; and (4) using a Random Forest classifier to determine if the overlap is significant enough to warrant academic credit restrictions.

The authors evaluated CourseGraph using a dataset from the Computer Science program at Eindhoven University of Technology (TU/e) and a holdout set of real-world exchange decisions from Lund University. The results demonstrate that CourseGraph outperforms simple threshold-based methods and zero-shot LLM prompting. While zero-shot LLMs showed high precision, they suffered from very low recall, making them too conservative for this application. In contrast, the Random Forest classifier achieved the best balance between precision and recall (F1 score of 0.74), proving effective at identifying content overlap consistent with human expert judgment.

CourseGraph provides a scalable, interpretable solution to a common bottleneck in higher education administration. By automating the identification of course overlaps, the tool can reduce the manual workload for program directors and academic advisors. Furthermore, because the model embeds individual course components (like learning outcomes) separately, it offers transparency, allowing administrators to see exactly which parts of a course contribute to an overlap prediction, thereby facilitating more informed and consistent decision-making.