Yesu Feng

Yesu Feng is a Netflix research scientist developing shared personalization foundation models that learn viewers’ preferences across movies, television, games, and other entertainment. His work replaces proliferating specialized recommenders with reusable systems that support discovery across products, interfaces, and content formats.

Feng earned a physics doctorate at Duke University in 2015, studying long-lived nuclear spin states. He subsequently worked on LinkedIn’s homepage-feed recommendations and, at Uber, on driver behavior and supply prediction before joining Netflix. There, he coauthored the company’s account of its foundation model for personalized recommendations, describing how centralized preference learning can reduce duplicated infrastructure while improving downstream recommendation systems.

• Modeling behavior as structured sequences. Feng applies autoregressive transformers to viewing histories, but treats each interaction as a complex event incorporating timing, device, context, content, and behavior. His interaction-event tokenization balances descriptive detail against context-window limits, while sliding-window training incorporates longer histories without increasing model input dimensions. Predicting multiple future behaviors helps capture enduring preferences instead of optimizing solely for the next click.

• Making shared recommendations production-ready. Feng combines learned item identifiers with semantic content embeddings to recommend unfamiliar titles before substantial viewing data exists. Downstream systems can reuse viewer representations and content embeddings, integrate pretrained model components, or apply fine-tuning and distillation under strict latency constraints. His explanation of Netflix’s unified recommendation architecture also outlines generative retrieval that incorporates diversity and business constraints when assembling collections.

• Grounding conversational discovery in collective behavior. Feng coauthored CRAG research on collaborative retrieval, combining conversational language models with collaborative-filtering signals to connect natural-language requests with actual audience preferences. The approach improves recommendation coverage, particularly for newly released movies.

• Personalizing both presentation and context. His subsequent research applies language-model post-training to personalized artwork, selecting different visual representations of the same title for different viewers. Related work on translating behavioral logs into language develops learned behavioral verbalization, converting interaction histories into cleaner, more useful context for language-model recommendations.