Keynote 20 - Train Once, Recommend Forever: Recsys for an Ever-Growing Catalog
Modern consumer products all share one trait: their catalog never stops growing. Streaming platforms, news feeds, marketplaces — new titles, articles, and listings arrive every day, and users expect to see them in their recommendations right away. But recommender systems learn exclusively from past user interactions, so an item that just arrived, with no interaction history yet, is invisible to them by design. The traditional fix is to retrain continuously — and that's exactly the trap: it couples what a system can recommend to when it was last trained. In this talk we follow the paradigm shift now reshaping recommender systems: decoupling a model's training from what it's allowed to recommend. Starting from the transformers that power modern discovery, we show why they cannot represent new items, how content embeddings almost solve the problem — and where they fall short — and how the approach we built at Albatross and run in production behind every feed, closes the gap: new items recommendable within minutes, retraining almost never, and quality above the ID-based baseline. Everything backed by production-scale experiments.
