Augmenting Netflix Search with In-Session Adapted Recommendations (RecSys 2022)
Paper: Augmenting Netflix Search with In-Session Adapted Recommendations, presented at RecSys 2022.
Overview
Netflix's product experience is powered by two major personalization surfaces: search, which serves members who arrive with a specific intent (a title, actor, or genre in mind), and recommendations, which personalize rows of content for members browsing without a concrete target. Historically these systems have been developed and operated largely as separate stacks.
This RecSys 2022 paper describes Netflix's work on augmenting search with recommendations that are adapted to the current session — using in-session signals (such as query intent and recent engagement) to condition recommendation models and blend their output into the search experience.
Motivation
- Many search sessions at Netflix exhibit ambiguous or evolving intent: members issue vague queries, refine them, or abandon search after unsatisfying results.
- Recommendation models excel at capturing member taste, but conventionally do not leverage the rich intent signals available within the active session.
- Bridging the two surfaces lets search present relevant titles the member may not have explicitly queried, improving discovery and reducing session abandonment.
- Adapt recommendation models using in-session context so that their output reflects the member's current intent rather than only long-term taste.
- Integrate the adapted recommendations into the search results page, complementing the core query-matching ranking.
- Serve the combined experience within Netflix's production search infrastructure.
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking and Recommendation
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
- Bridging Language and Items for Retrieval and Recommendation (2024)
- Data-efficient Fine-tuning for LLM-based Recommendation (SIGIR 2024)
- DiffKG: Knowledge Graph Diffusion Model for Recommendation (WSDM 2024)
Approach (as described at a high level)
Evaluation
Consistent with industry practice at Netflix, the system was assessed through online A/B experiments measuring member engagement and satisfaction-related metrics, alongside offline model quality checks. Exact metric values and experimental configurations should be taken from the original paper.
Takeaways
1. Search and recommendation are complementary: session-adapted recommendations can meaningfully augment query-driven results, especially for ambiguous intent. 2. In-session signals are valuable: short-term behavior within a session is a strong personalization signal distinct from long-term member profiles. 3. Production integration matters: unifying two major surfaces requires careful system design and rigorous online experimentation.
> Note: The source forum post is a templated overview with limited technical detail. For precise methods, results, and figures, refer to the original publication: DOI 10.1145/3523227.3547407.