Workflow · Production

FM-Intent: Predicting User Session Intent with Hierarchical Multi-Task Learning at Netflix

The problem

Netflix's foundation model focused on next-item prediction but lacked the ability to capture or leverage underlying user session intent, and existing intent prediction approaches did not establish a hierarchical relationship between intent and item prediction tasks.

First attempt

Prior approaches to intent prediction used simple multi-task learning without a hierarchical structure, and most baseline models either could not predict user intent or could not incorporate intent predictions into next-item recommendations.

Workflow diagram · grounded in source
1
Input feature construction
integration
“The input feature for each interaction combines categorical embeddings and numerical features, creating a comprehensive representation of user behavior.”
2
Intent prediction via Transformer
ai_action
“The intent prediction component processes the input feature sequence through a Transformer encoder and generates predictions for multiple intent signals. The Transformer encoder effectively models the long-term interest of users through …”
3
Attention-based intent aggregation
ai_action
“This approach generates a comprehensive intent embedding that captures the relative importance of different intent signals for each user”
4
Hierarchical next-item prediction
ai_action
“FM-Intent employs hierarchical multi-task learning where intent predictions are conducted first, and their results are used as input features for the next-item prediction task.”
5
Production ecosystem integration
output
“FM-Intent has been successfully integrated into Netflix's recommendation ecosystem”
Reported outcome

FM-Intent achieves a statistically significant 7.4% improvement in next-item prediction accuracy over the best baseline and has been successfully integrated into Netflix's recommendation ecosystem.

Reported metrics
Next-item prediction accuracy improvement7.4%
Reported stack
Transformer encoderK-means++
Source
https://netflixtechblog.com/fm-intent-predicting-user-session-intent-with-hierarchical-multi-task-learning-94c75e18f4b8
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

FM-Intent achieves a statistically significant 7.4% improvement in next-item prediction accuracy over the best baseline and has been successfully integrated into Netflix's recommendation ecosystem.

What tools did this team use?

Transformer encoder, K-means++.

What results were reported?

Next-item prediction accuracy improvement: 7.4% (source-reported, not independently verified).

What failed first in this deployment?

Prior approaches to intent prediction used simple multi-task learning without a hierarchical structure, and most baseline models either could not predict user intent or could not incorporate intent predictions into ne…

How is this workflow AI workflow structured?

Input feature construction → Intent prediction via Transformer → Attention-based intent aggregation → Hierarchical next-item prediction → Production ecosystem integration.