Novel approach on user match making using LLM generated vectors from YouTube's recommendation algorithm
Shailaja V. Pede, Atharva Galne, Vaibhav Gangurde, Utkarsh Patil, Sarthak Kshirsagar · 2025
Understanding human interests is complex, making them difficult to quantify. Traditional user profiling and matchmaking often depends on self-reported interests, which may be incomplete or biased. This research explores an automated approach to user profiling by analysing YouTube recommendations, which are shaped by deep learning-based content suggestion algorithms. Our method involves collecting video recommendations and subscription data, categorising them into a structured interest taxonomy, and representing users as vectorised interest profiles. To enhance classification, large language models (LLMs) extract key interest categories from video metadata. Additionally, we examine co-subscriber networks to improve user matchmaking based on shared content consumption. Experimental results suggest that analysing recommendation data offers a more comprehensive and accurate view of user interests than manual methods, leading to better profiling and matchmaking outcomes.