Bypassing the Popularity Bias

In this episode, we speak with Václav Blahut, a machine learning researcher at Seznam.cz, about tackling popularity bias in recommender systems. Václav explains inverse recommendation—finding the right users for niche content rather than items for users—and how Seznam.cz repurposed their two-tower retrieval model to implement this approach. He discusses real-world metrics like "bottom 50% share," cold start challenges, and how combining multiple recommendation strategies with personalized content mixes yields better results. Václav holds a Master's degree in AI and NLP from Masaryk University and has over six years of experience in recommender systems. This episode discusses the paper: "Bypassing the Popularity Bias: Repurposing Models for Better Long-Tail Recommendations"

Guest

Václav Blahut: Václav mastered AI & NLP course at FI MUNI and is now teaching machines how to recommend articles and other content to Seznam.cz users. He's been doing it for more than six years now. Currently, he is trying to figure out how to improve content-based recommendation of external content with sparse feedback and how to improve recommendations in general.