Please note: This master’s thesis presentation will take place online.
Doaa Elfayoumi, Master’s candidate
David R. Cheriton School of Computer Science
Supervisor: Professor Charles Clarke
Choosing university courses can be overwhelming for students, especially in large institutions where hundreds of options are offered each term. Motivated by these observations, I envisioned developing a system that could help students identify educational pathways aligned with their goals and aspirations. By leveraging the University of Waterloo’s course and student data, this research aims to design a recommendation system that suggests courses based on students’ interests and desired career outcomes.
To address several challenges, including the lack of aggregate historical enrollment data, this thesis presents a new AI-based course recommendation system designed specifically for the University of Waterloo. The system combines a knowledge graph, semantic similarity techniques, and a custom \textit{interestingness} score to recommend courses that are not only eligible but also meaningful and relevant to the student.
The knowledge graph models relationships between courses, programs, prerequisites, terms, and subjects, allowing the system to reason about eligibility in a transparent way. To capture course similarity, I compare two embedding models, sBERT and OpenAI embeddings, and evaluate them by examining their similarity matrices and observing how subjects naturally cluster together. Based on these results, sBERT was selected for the final system due to its strong performance and ability to run locally without API costs. The system also integrates a personalized scoring method that considers prerequisite depth, program requirements, and students’ enrollment history, helping surface courses students might not otherwise discover.
This thesis also explores how Large Language Models (LLMs), such as ChatGPT, can support recommendation tasks. Recent studies show that LLMs can enhance existing recommenders through feature extraction, prompt-based ranking, explanation generation, and improved handling of cold-start cases. In this project, LLMs are used as a complementary tool, primarily for generating code from text that represents course prerequisites. The resulting system embodies a hybrid AI approach that blends LLM-driven semantic understanding with knowledge-graph-based reasoning to deliver more intelligent, context-aware recommendations.
Overall, this work demonstrates a practical and scalable approach to course recommendation. By combining structured academic data with semantic similarity and AI-driven reasoning, this research shows how natural language processing, large language models, and knowledge-graph reasoning can come together to create an adaptable and effective framework for educational recommendation systems.