Netflix's Big Bet: One model to rule recommentdations (AI)

I saw something interesting yesterday. speech Yesu Feng (Yesu Feng) from Netflix about using generative foundation models for personal recommendations. By the way, before joining Netflix, Yesu Feng worked at LinkedIn on the news feed and at Uber on the optimization of the marketplace. I have identified the following key ideas:

1. A single foundation model for the entire recommendation system Instead of many specialized models (for different pages, genres, content formats) Netflix is moving to one “basic” autogressive transformer model, capable of covering all use cases. 2. Scaling as the main driver of quality They came up with two fairly logical hypotheses. As the volume of data and model parameters increase, personalization improves according to the same laws of scaling as LLM. Integrating this model into all subsystems creates synergies and accelerates innovation. 3. Features of data and learning As a result, the model was multilevel. At the basic level, the guys did event representation: when, where (locale & device & canvas), what (action type & entity & duration related) Then there was the embedding/feature transformation level - here it was necessary to combine id embedding and additional semantic embeddings to solve the problem of a cold start. (For example, new content) The next level contained transformer/attention - hidden state layers were used as user representation, it was necessary to guarantee the stability of the users' representation, it was also necessary to be able to clearly adapt to different user goals, and it was necessary to understand how to aggregate different levels and different sequences to obtain this representation. At the top level was the objective loss function, which was complicated because the LLM output had multiple sequences. (sequences) Multitasking at the level of the loss function (prediction of the type of action, duration of the session, device, etc.) -- main objective: entity id -- auxiliary objectives: action types, entity metadate, duration, device, time -- reward, weight & mask 4. Lessons from the LLM world When building an LLM model, the guys stuffed a number of bumps and they shared them. Multi-tone prediction to increase resistance to time shifts and focus on long-term behavior. Multi-layer presentation and self-distillation for the stability of the user's embeddings. Processing long context windows through progressive length increase and sparse attention. 5. Application scenarios of the foundation model They came up with the following scenarios. Embedding as a subgraph in a downstream model. Exporting and updating user emblems and content in central storage. Retraining or distillation for highly specialized tasks with stringent delay requirements. 6. Results and conclusions Scaling the model from tens of millions to billions of parameters confirms the patterns of quality growth as data increases. The introduction of a single model has led to significant A/B gains and infrastructure consolidation: faster development of new features and reduced duplication of effort. Netflix’s core “bet” was fully justified: the foundation model turned out to be a scalable and flexible solution for personalizing recommendations.

This approach is planned to be developed in the following directions: Universal representations for heterogeneous entities (Videos, games, live broadcasts, etc.). Generative selection of collections using multi-step decoding. Quick adaptation through prompt-tuning and soft tokens for rapid change of model goals.

#AI #Engineering #ML #Architecture #Software #Data