Background: Recommender systems pervade our everyday lives from social networks to media to content streaming apps. While we like being recommended just the right item, the number of items in the catalogs have increased manifold. This presents a unique opportunity for the AI systems to get better at retrieval as well as ranking tasks. While these systems know a lot of things about existing users, they know very little about cold-start users or users just starting on a platform. How do we personalize content better for new users? This presents a unique challenge of being cold-start aware in recommender systems. Finally, we need to do this at scale of several hundred million users across dozens of markets.
Talk: In this talk I will present our journey on transfer learning in recommender systems and how we have enabled a culture of leveraging foundational learning in the full recommendation model stack including use of embeddings for both retrieval and ranking. This includes item as well as user embeddings. This talk will present how audience members can leverage transfer learning in their own stack – and not just for recommendation tasks but just about any predictive tasks. Some experiences on doing this at scale of millions of users will also be shared. The talk will present our journey in building robust evaluation systems that work for both new and existing users. I’ll also cover some of the recent learnings from leveraging the modern AI stack with Large Language Models (LLMs) in recommender tasks.