Enhancing Netflix Recommendations with FM-Intent: Predicting User Session Intent

Understanding Netflix’s FM-Intent: A Revolution in Recommendation Systems

Netflix’s recommendation system has long been a cornerstone of its user experience, helping members discover content that matches their preferences. Now, the streaming giant has taken a significant leap forward with FM-Intent, a sophisticated model that predicts underlying user intentions during viewing sessions.

The Challenge of Understanding User Intent

Traditional recommendation systems focus primarily on predicting the next item a user might like based on their viewing history. However, these systems often lack understanding of why users engage with content in specific ways during different sessions. A user might be in “discovery mode” one evening, looking to find new shows, while in another session they might simply want to continue watching a series they’ve started.

What Makes FM-Intent Different?

FM-Intent employs a hierarchical multi-task learning approach with three key components:

1. Input Feature Sequence Formation – Combines interaction metadata to create comprehensive representations of user behavior
2. User Intent Prediction – Utilizes Transformer encoders to model long-term interests and predict multiple intent signals
3. Next-Item Prediction – Leverages predicted user intent to improve recommendation accuracy

The model’s innovation lies in its ability to first predict user intent and then use that prediction to inform content recommendations, creating a more coherent and effective recommendation pipeline.

Types of User Intent Captured

FM-Intent captures several dimensions of user intent:

  • Action Type – Whether users want to discover new content or continue watching something
  • Genre Preference – The specific content categories users prefer in a given session
  • Movie/Show Type – Whether users are looking for a movie or a TV series
  • Time-since-release – If users prefer newly released, recent, or catalog content

Performance Improvements

Testing shows that FM-Intent delivers a significant 7.4% improvement in next-item prediction accuracy compared to the best baseline model. This represents a substantial enhancement in Netflix’s ability to recommend relevant content.

Real-World Applications

The intent prediction capabilities of FM-Intent enable several practical applications:

  • Personalized UI optimization based on whether users are in discovery or continue-watching mode
  • Enhanced user analytics for content acquisition and production decisions
  • Improved search functionality that prioritizes results based on current session intent
  • More relevant recommendations through better understanding of viewing patterns

User Clustering Insights

One fascinating aspect of FM-Intent is its ability to generate meaningful user intent embeddings for clustering. The model identifies distinct viewer segments such as content discoverers, series continuers, genre enthusiasts, and users with specific viewing patterns like rewatchers.

The Future of Personalized Streaming

By understanding not just what users might watch next but the underlying intentions driving their behavior, Netflix can provide a more nuanced, personalized experience that adapts to the shifting needs of viewers from session to session.

This advancement represents the evolution of recommendation systems from simple “you might also like” suggestions to sophisticated platforms that understand the context and purpose behind each viewing session.

For streaming services competing in an increasingly crowded marketplace, this level of personalization could be a crucial differentiator that keeps subscribers engaged and satisfied with the service.

Visit the Netflix Tech Blog for more detailed information on FM-Intent and hierarchical multi-task learning