Keynote Xavier Amatriain

The keynote traces a joint history of AI and recommender systems, starting with MovieLens (1997), the Netflix Prize (2006–2009), and the transition from rating prediction and RMSE toward ranking, page optimization, and implicit feedback. Amatriain reviews matrix factorization, RBMs, and early Netflix production lessons, then the rise of deep learning and two-tower models. He highlights the enduring importance of UX, domain knowledge, and offline/online evaluation, noting diversity’s causal link to long-term satisfaction. The talk surveys transformers and LLMs for preference understanding, generative retrieval, and semantic features, with live demos of Gemini-based recommendation elicitation and an event-finding agent. Looking ahead, he discusses agents, world models (e.g., Genie 3), continuous user memory, RAG plus fine-tuning, and the prospect of personalized content generation, along with open cultural and evaluation challenges.