Foundation Model for Large-Scale Personalized Recommendation

netflix · Media & Entertainment · 2025 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: personalized recommendation foundation models

Netflix's personalized recommender system had become costly to maintain because many specialized models were independently trained on common data, making innovation transfer difficult while serving-latency constraints limited how much user history each model could use.

Constraints:

• many specialized recommendation models served distinct product needs such as Continue Watching and Today's Top Picks
• member interactions span hundreds of billions of events but require careful tokenization and quality filtering
• active users can have thousands of events while recommender inference must run at millisecond-level latency
• new titles need recommendations before user interaction data exists
• downstream consumers need stable embeddings across foundation-model retraining and redeployment

Architecture

Netflix built a recommendation foundation model inspired by LLMs. It tokenizes member interactions into meaningful sequence events, trains an autoregressive next-token model with recommendation-specific objectives, uses sparse attention and sliding-window sampling to learn from long histories, handles cold-start with incremental training and metadata-based title embeddings, and exposes the model through direct prediction, batch embeddings, candidate generation, and fine-tuning for downstream recommendation applications.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  user-interaction-history[User interaction history]
  interaction-tokenizer[Interaction tokenizer]
  autoregressive-recommendation-model[Autoregressive recommendation model]
  sparse-attention-and-sampling[Sparse attention and sampling]
  cold-start-embedding-mixer[Cold-start embedding mixer]
  embedding-stabilization-layer[Embedding stabilization layer]
  downstream-recommendation-applications[Downstream recommendation applications]
  user-interaction-history --> interaction-tokenizer
  interaction-tokenizer --> autoregressive-recommendation-model
  autoregressive-recommendation-model --> sparse-attention-and-sampling
  sparse-attention-and-sampling --> autoregressive-recommendation-model
  autoregressive-recommendation-model --> cold-start-embedding-mixer
  cold-start-embedding-mixer --> autoregressive-recommendation-model
  cold-start-embedding-mixer --> embedding-stabilization-layer
  embedding-stabilization-layer --> downstream-recommendation-applications
  downstream-recommendation-applications --> embedding-stabilization-layer
  downstream-recommendation-applications --> autoregressive-recommendation-model

Components

Data flow

Raw member activity is tokenized by merging meaningful interactions and retaining details such as duration, locale, time, device, title ID, genre, and release country. Training samples overlapping windows from long histories and uses autoregressive objectives plus auxiliary targets to learn member preferences. New-title metadata and ID embeddings are mixed based on entity age. Downstream systems then consume predictions, stable embeddings, candidate-generation features, or fine-tuned subgraphs.

Why this architecture

Netflix centralized preference learning because maintaining many independently trained recommendation models was costly and made it hard to transfer innovations. The design borrows LLM scaling and semi-supervised next-token learning, but adapts it for recommender-specific constraints such as millisecond latency, unequal interaction importance, new-title cold start, and stable downstream embeddings.

Models used: Netflix recommendation foundation model · autoregressive next-token prediction model · sparse attention mechanisms · KV caching

Scale

• Netflix had over 300 million users at the end of 2024.
  “With over 300 million users at the end of 2024, this translates into hundreds of billions of interactions”article, stated
• The interaction corpus spans hundreds of billions of events, comparable to LLM token volume.
  “hundreds of billions of interactions — an immense dataset comparable in scale to the token volume of large language models (LLMs).”article, stated
• Active member histories can span thousands of events.
  “interaction histories from active users can span thousands of events, exceeding the capacity of transformer models with standard self attention layers.”article, stated
• Recommendation serving requires millisecond-level latency, unlike LLM applications that may tolerate seconds.
  “these services typically require millisecond-level latency. This constraint is more stringent than what is typical in LLM applications”article, stated
• Sparse attention extends the model context to several hundred events.
  “the model can extend its context window to several hundred events while maintaining computational efficiency.”article, stated
• A 5-minute trailer play and a 2-hour movie watch are explicitly treated as different-strength interactions.
  “a 5-minute trailer play should not carry the same weight as a 2-hour full movie watch.”article, stated

Alternatives considered (roads not taken)

maintain numerous independently trained specialized recommendation models

Rejected because: The model fleet had become costly and made innovation transfer difficult despite shared data sources.

extract features only from brief recent interaction windows

Rejected because: Short windows miss comprehensive member history, but full histories exceed practical latency and training constraints.

treat every target interaction equally during pretraining

Rejected because: Recommendation events differ in importance, and long-term satisfaction cannot be aligned by weighting every interaction equally.

use only item ID embeddings for new titles

Rejected because: New and pre-launch titles may have little or no interaction data, so metadata must help estimate preferences.

expose raw embedding spaces that change after each retraining

Rejected because: Arbitrary, incompatible embedding dimensions create bugs and force downstream consumers to adapt after every retrain.

Patterns

retrieval-ranking (personalization)

Netflix uses a foundation model to generate predictions, embeddings, and candidate-generation features for personalized recommendation surfaces.

fine-tuning-adaptation (model-adaptation)

The foundation model can be integrated directly or through subgraphs and fine-tuned with application-specific data using less data and compute.

eval-driven-development (model-quality)

Netflix treats scaling as dependent on robust evaluation that can distinguish model improvements and identify where the recommender should improve.

Tradeoffs

Centralize preference learning in a foundation model.

Chose a unified recommendation foundation model over many independently trained specialized models.

Rationale: Centralization reduces maintenance cost and makes shared preference learning reusable across downstream models.

Balance sequence compression with retained interaction detail.

Chose interaction tokenization that merges adjacent actions while preserving key fields over raw fine-grained event sequences or overly lossy tokens.

Rationale: The model needs long histories but cannot afford unbounded sequence length or dropped behavioral signals.

Use multi-token and auxiliary objectives instead of plain next-item prediction alone.

Chose multi-token prediction plus auxiliary targets such as genre over equal-weight single-token item prediction.

Rationale: The modified objective captures longer-term dependencies, reduces overfitting on noisy item IDs, and narrows candidate lists.

Mix metadata and ID embeddings by entity age.

Chose age-aware attention between metadata-based and ID-based title embeddings over simply summing metadata and ID embeddings.

Rationale: New titles should rely more on metadata, while established titles can rely more on learned engagement-based IDs.

Expose stable transformed embeddings to downstream systems.

Chose orthogonal low-rank transformation over forcing downstream consumers to adapt to each retrained embedding space.

Rationale: Stabilized dimensions reduce bugs and invalid assumptions when the foundation model is retrained and redeployed.

Gotchas (what bit them)

Specialized recommender fleets create maintenance and transfer costs.

Root cause: Models are independently trained even when they use common data sources, so each innovation must be moved model by model.

Resolution: Centralize member preference learning in a foundation model that can distribute weights, embeddings, and fine-tuning capabilities.

Tokenizing recommendation behavior can lose either detail or tractability.

Root cause: Overly lossy tokenization drops valuable signals, while too-granular sequences exceed processing-time and memory limits.

Resolution: Merge adjacent actions into meaningful tokens while preserving important information such as watch duration and engagement types.

Recommender context windows are constrained by serving latency rather than only model capacity.

Root cause: Online recommender services require millisecond-level inference, making long histories harder to use than in typical LLM apps.

Resolution: Use sparse attention, sliding-window sampling, and KV caching to learn from long histories while maintaining low latency.

New titles have no interaction history.

Root cause: Recommendation models must estimate preferences for newly launched or pre-launch titles before members have engaged with them.

Resolution: Warm-start new embeddings and combine metadata-based embeddings with ID-based embeddings using an age-aware attention mechanism.

Embedding spaces change across retraining runs.

Root cause: The user/item embedding dimensions are arbitrary, uninterpretable, and incompatible across model versions.

Resolution: Apply an orthogonal low-rank transformation to stabilize embedding dimensions for downstream consumers.

Evaluation

Netflix evaluates the foundation model through downstream integrations, scaling-law experiments, model/data/context scaling, prediction quality, embedding usefulness, and evaluation methods that can distinguish performance and identify improvement areas.

Metrics: over 300 million users at end of 2024 · hundreds of billions of interactions · context windows extended to several hundred events · active histories spanning thousands of events · millisecond-level serving latency · scaling-law improvements as data and model size increase · promising downstream integration results

Outcomes

• Netflix moved toward a unified, data-centric recommendation foundation model instead of many specialized models.
• The model supports direct prediction, embeddings for downstream offline and online applications, candidate generation, and fine-tuning.
• Experiments showed scaling laws apply to the recommendation foundation model as data and model size increase.
• Cold-start handling supports newly launched or pre-launch titles by combining metadata and ID embeddings.
• Embedding stabilization reduces downstream breakage across retraining and redeployment.

The source reports over 300 million users, hundreds of billions of interactions, active histories of thousands of events, several-hundred-event context via sparse attention, millisecond-serving constraints, and promising downstream integrations.