This paper introduces a novel course recommendation system using LLMs and RAG to help university students discover relevant courses based on natural language queries .
-----
https://arxiv.org/abs/2412.19312v2
Original Problem 🎯:
Students at large universities struggle to navigate thousands of course options each term. Traditional recommender systems lack personalization, interactivity, and fail for new students without historical data .
-----
Solution in this Paper 🔧:
→ The system uses a two-stage approach combining LLMs with embedding-based similarity search
→ First, it generates an "ideal" course description from the user's natural language query using GPT-3.5-turbo
→ This description is converted into a search vector using embeddings to find similar actual courses
→ The system then uses GPT-4 to analyze the top 50 most relevant courses and provide 10 final recommendations with explanations and confidence ratings
-----
Key Insights 🔍:
→ The embedding space effectively captures meaningful relationships between academic subjects
→ Top similarity-ranked courses appear in final recommendations 85% of the time
→ The system shows no significant demographic biases in core recommendations
→ Recommendations maintain consistency across diverse query types
-----
Results 📊:
→ Context retrieval time: 2-5 seconds
→ Total recommendation generation time: 9-13 seconds
→ Top 12 similarity ranks contribute to 60% of recommended courses
→ System maintains recommendation quality across broad and specialized queries
Share this post