Keynote 13: Train Once, Recommend Forever: Recsys for an Ever-Growing Catalog
Modern consumer products all share one trait - their catalog never stops growing. Think streaming platforms, marketplaces, news feeds: new shows, new listings, new articles arrive every day, and users expect to see them in their recommendations right away. But recommendation systems learn exclusively from past user interactions, so anything that just arrived, with no interaction history yet, is invisible to them by design. The traditional fix is to retrain continuously - but that's exactly the trap: it couples what a system can recommend to when it was last trained. New items go unrecommended, and hidden gems stay invisible. In this talk, we discuss the paradigm shift now sweeping through recommender-systems research: decoupling the training of a recommender from what it's allowed to recommend. We explore the techniques that let Albatross and tech giants like YouTube, TikTok, and Amazon serve high-quality recommendations from ever-growing, billion-item catalogs in milliseconds. Drawing on production-scale experiments, we take a tour through the architectures making this possible: from the transformer that's the workhorse of modern discovery systems, through the generative-retrieval wave currently exciting the field, to the frontier approach we've built at Albatross - backed by comparisons and hands-on examples.
