Hierarchical Multi-Task Learning for Intent Prediction in Recommender Systems

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: streaming media recommender systems

Netflix wanted its recommendation foundation model to understand the intent behind a viewing session, not only predict the next item, so recommendations, search, UI layout, and analytics could respond to whether a member was discovering, continuing, exploring genres, or seeking a movie versus a show.

Constraints:

• latent session intent is not directly observable
• intent can vary across action type, genre preference, movie/show preference, and time-since-release preference
• existing intent-prediction approaches typically add heads without a hierarchical relationship to item recommendation
• the model must preserve next-item recommendation quality while adding intent prediction
• the complex hierarchical architecture uses a smaller training dataset than Netflix's production foundation model for offline comparison
• some baseline models cannot predict user intent or cannot feed intent back into item prediction
• the extracted intent embedding must be useful for production ecosystem applications beyond the standalone model

Architecture

Netflix built FM-Intent as an extension of its recommendation foundation model. It constructs rich input feature sequences by combining interaction metadata, encodes user history with a Transformer encoder, predicts multiple intent signals through task-specific fully connected layers, aggregates those intent predictions with attention to form a comprehensive intent embedding, and feeds that embedding into the next-item prediction task. The hierarchy is intentional: intent is predicted first and then used as an input feature for next-item prediction, rather than being learned as a disconnected auxiliary head. Offline experiments compare FM-Intent against sequential recommendation baselines and a production-model variant, while downstream integration exposes intent embeddings and predictions for personalized UI optimization, analytics, enhanced recommendation signals, and search optimization.

Diagram confidence: reconstructed (reconstructed from prose).

flowchart LR
  A["Netflix interaction metadata"] --> B["Input feature sequence"]
  B --> C["Transformer encoder"]
  C --> D["Intent prediction heads"]
  D --> E["Attention-based intent aggregation"]
  B --> F["Next-item prediction head"]
  E --> F
  F --> G["Netflix recommendation ecosystem"]
  E --> H["User clustering analysis"]
  I["Offline evaluation suite"] --> D
  I --> F
  G --> J["Personalized UI optimization"]
  G --> K["Analytics and user understanding"]
  G --> L["Enhanced recommendation signals"]
  G --> M["Search optimization"]

Components

Data flow

Netflix interaction metadata is combined into input feature sequences with categorical embeddings and numerical features. A Transformer encoder models long-term user interest through multi-head attention, and task-specific heads predict intent dimensions such as action type, genre preference, movie/show type, and time-since-release. Attention-based aggregation combines the individual intent predictions into an intent embedding. The next-item prediction model then combines that intent embedding with the input features to produce recommendations. Intent embeddings can also be clustered or passed into other recommendation, UI, analytics, and search systems.

Why this architecture

Netflix chose a hierarchical multi-task design because next-item prediction alone cannot represent why the user is browsing, and conventional auxiliary intent heads do not make intent directly inform the recommendation. Predicting intent first and feeding the resulting embedding into item prediction creates a coherent pipeline where recommendations can be conditioned on session-level purpose. Attention-based aggregation also makes the multiple intent signals useful as a single embedding for personalization and explanation.

Models used: Netflix recommendation foundation model · FM-Intent · Transformer encoder · LSTM baseline · GRU baseline · Transformer baseline · TransAct baseline · FM-Intent-V0

Scale

• The work targets Netflix's production recommendation ecosystem.
  “FM-Intent has been successfully integrated into Netflix’s recommendation ecosystem”article, stated
• Offline experiments used sampled Netflix user engagement data.
  “We conducted comprehensive offline experiments on sampled Netflix user engagement data to evaluate FM-Intent’s performance.”article, stated
• FM-Intent improved next-item prediction accuracy by 7.4% versus TransAct in offline experiments.
  “FM-Intent demonstrates statistically significant improvement of 7.4% in next-item prediction accuracy compared to the best baseline (TransAct).”article, stated
• User intent clustering was shown with 10 K-means++ clusters.
  “Figure 3 visualizes 10 distinct clusters identified through K-means++ clustering.”article, stated
• The offline comparison trained FM-Intent on a much smaller dataset than the actual production foundation model.
  “FM-Intent uses a much smaller dataset for training compared to the FM production model due to its complex hierarchical prediction architecture.”article, stated

Alternatives considered (roads not taken)

Use the recommendation foundation model for next-item prediction only

Rejected because: The prior foundation model understood interaction histories but lacked explicit prediction and use of underlying session intent.

Add auxiliary intent-prediction heads without hierarchy

Rejected because: Netflix argues that conventional simple multi-task learning does not establish a hierarchy where intent predictions directly inform item recommendations.

Use baseline sequential recommendation models

Rejected because: The tested baselines either could not predict user intent or could not incorporate intent predictions into next-item recommendations.

Use the actual larger production foundation model for the offline baseline

Rejected because: For fair comparison, Netflix used FM-Intent-V0 trained with a smaller dataset, while noting that the actual production model has much more training data.

Patterns

fine-tuning-adaptation (model adaptation)

Netflix adapted its recommendation foundation model by adding hierarchical intent prediction and an intent-informed next-item prediction path for session-level personalization.

retrieval-ranking (ranking)

Predicted session intent is used to improve next-item recommendation ranking and can also prioritize search results based on the user's current session intent.

eval-driven-development (evaluation)

The model was validated through offline experiments against state-of-the-art sequential recommendation baselines and a production-model variant before production ecosystem integration.

structured-output-contract (model interface)

The model turns latent session behavior into explicit intent dimensions such as action type, genre preference, movie/show type, and time-since-release, giving downstream systems structured signals to consume.

Tradeoffs

Use hierarchical multi-task learning

Chose intent predictions feeding next-item prediction over flat auxiliary heads attached to a next-item model.

Rationale: The hierarchy makes the predicted intent a functional input to recommendations rather than a parallel label that may not improve item ranking.

Aggregate multiple intent predictions with attention

Chose attention-based aggregation into a comprehensive intent embedding over treating each intent head independently.

Rationale: Attention lets the model represent the relative importance of different intent signals for each user and provides a single embedding for downstream personalization and explanation.

Use a smaller training dataset for FM-Intent experiments

Chose sampled user engagement data and a fair smaller-dataset baseline over direct comparison against the full larger production foundation model.

Rationale: The hierarchy is complex enough that Netflix trained on less data, so the baseline was constrained similarly to isolate the modeling improvement.

Use intent embeddings beyond direct recommendation

Chose downstream UI, analytics, feature, and search applications over limiting FM-Intent to next-item prediction only.

Rationale: The intent embedding exposes a reusable representation of session purpose, allowing other systems to optimize UI, analyze audiences, enhance models, and prioritize search.

Gotchas (what bit them)

Auxiliary intent heads alone do not guarantee intent-informed recommendations.

Root cause: Simple multi-task learning can predict intent as a side task without making those predictions influence next-item recommendation.

Resolution: FM-Intent imposes a hierarchy where intent predictions are computed first and then used as input features for the next-item prediction task.

Intent is latent and only observable through proxies.

Root cause: The model cannot directly observe why a user is browsing; it must infer intent from action type, genre preference, content type, release recency, and other interaction metadata.

Resolution: FM-Intent predicts multiple intent dimensions from implicit interaction signals and aggregates them into a comprehensive embedding.

Complex hierarchical prediction increases training cost enough to affect offline comparison design.

Root cause: FM-Intent's architecture is more complex than the production foundation model path.

Resolution: Netflix trained FM-Intent on a smaller sampled dataset and compared against a production-model variant trained on similarly constrained data for fairness.

Many baselines cannot use intent as both prediction target and recommendation feature.

Root cause: Some sequential recommendation models are not designed to predict intent, and others cannot feed intent predictions back into item ranking.

Resolution: FM-Intent uses hierarchical multi-task learning so intent prediction is part of the path to next-item prediction.

Evaluation

Netflix evaluated FM-Intent with offline experiments on sampled user engagement data, comparing next-item and next-intent prediction against state-of-the-art sequential recommendation baselines and FM-Intent-V0. The evaluation reports relative percentage improvements versus TransAct and qualitative K-means++ clustering over intent embeddings.

Metrics: next-item prediction accuracy · next-intent prediction accuracy · relative percentage improvements versus TransAct · statistically significant 7.4% improvement in next-item prediction accuracy · K-means++ user clustering with K=10

Outcomes

• FM-Intent improved next-item prediction accuracy by 7.4% versus TransAct in offline experiments.
• The model generated meaningful intent embeddings that formed 10 distinct K-means++ clusters.
• FM-Intent was integrated into Netflix's recommendation ecosystem.
• Predicted intent can support personalized UI optimization, analytics, enhanced recommendation signals, and search optimization.

The paper describes offline experiments on sampled Netflix user engagement data, a model integrated into Netflix's recommendation ecosystem, and downstream uses in homepage layout, analytics, other recommendation models, and real-time search optimization.