Back office ops · Production

Netflix builds a foundation model for personalized recommendation to centralize member preference learning at scale

The problem

Netflix operated a large set of specialized recommendation models that were costly to maintain and difficult to share innovations across, and most were confined to brief temporal windows of user interaction history due to serving latency and training cost constraints.

Workflow diagram · grounded in source
1
User interaction events collected
trigger
“user engagement spans a wide spectrum, from casual browsing to committed movie watching. With over 300 million users at the end of 2024, this translates into hundreds of billions of interactions”
2
Interaction tokenization
ai_action
“Tokenization helps define what constitutes a meaningful "token" in a sequence. Drawing an analogy to Byte Pair Encoding (BPE) in NLP, we can think of tokenization as merging adjacent actions to form new, higher-level tokens”
3
Foundation model training
ai_action
“our default approach employs the autoregressive next-token prediction objective, similar to GPT. This strategy effectively leverages the vast scale of unlabeled user interaction data”
4
Cold-start embedding initialization
ai_action
“new title embeddings can be initialized by adding slight random noise to existing average embeddings or by using a weighted combination of similar titles' embeddings based on metadata”
5
Embedding generation and storage
output
“The model generates valuable embeddings for members and entities like videos, games, and genres. These embeddings are calculated in batch jobs and stored for use in both offline and online applications.”
6
Downstream fine-tuning and integration
integration
“Users can integrate the full model or subgraphs into their own models, fine-tuning them with less data and computational power”
Reported outcome

The foundation model enables centralized member preference learning distributable to downstream models via embeddings or fine-tuning, with promising results from downstream integrations and consistent improvements as data and model scale increase.

Reported metrics
Netflix user baseover 300 million users
User interaction data scalehundreds of billions of interactions
Downstream integration resultspromising results
Model performance improvement with scaleconsistent improvements
Reported stack
transformer modelsKV cachingsparse attentionSliding Window Sampling
Source
https://netflixtechblog.com/foundation-model-for-personalized-recommendation-1a0bd8e02d39
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

The foundation model enables centralized member preference learning distributable to downstream models via embeddings or fine-tuning, with promising results from downstream integrations and consistent improvements as…

What tools did this team use?

transformer models, KV caching, sparse attention, Sliding Window Sampling.

What results were reported?

Netflix user base: over 300 million users; User interaction data scale: hundreds of billions of interactions; Downstream integration results: promising results; Model performance improvement with scale: consistent improvements (source-reported, not independently verified).

How is this back office ops AI workflow structured?

User interaction events collected → Interaction tokenization → Foundation model training → Cold-start embedding initialization → Embedding generation and storage → Downstream fine-tuning and integration.

Related back office ops cases