Augmenting Netflix Search with In-Session Adapted Recommendations (RecSys 2022)
- Source: ACM Digital Library — DOI 10.1145/3523227.3547407
- Venue: RecSys 2022
- Affiliation: Netflix
- Category: Recommender Engines / Search
- A production approach to augmenting search results pages with in-session adapted recommendations, rather than treating search and recommendation as disjoint surfaces.
- Integration of session-level behavioral signals alongside query semantics to tailor recommended candidates.
- An online experimentation framework at Netflix scale to evaluate engagement impact of the hybrid search-recommendation experience.
- Engineering lessons on candidate generation, ranking, and merging heterogeneous result sources in one page.
- Setting: online A/B experiments on the Netflix production search experience.
- Metrics: member engagement measures such as play rate, successful discovery, and session satisfaction, complemented by offline relevance checks.
- Baselines: the existing search-only results page.
- 360Brew: A Decoder-only Foundation Model for Personalized Ranking
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers
- Bridging Language and Items for Retrieval and Recommendation (Mar 2024)
- Data-efficient Fine-tuning for LLM-based Recommendation, SIGIR 2024
- DiffKG: Knowledge Graph Diffusion Model for Recommendation, WSDM 2024
- EAGER-LLM: Enhancing Large Language Models as Recommenders through Exogenous...
- Original paper: *Augmenting Netflix Search with In-Session Adapted Recommendations*, RecSys 2022. https://dl.acm.org/doi/abs/10.1145/3523227.3547407
One-line summary
Netflix augments its search results pages with recommendations that are adapted within the session, blending query relevance with dynamic personalization.
Background and Motivation
Search and recommendation have traditionally been treated as separate retrieval paradigms: search matches an explicit query against a catalog, while recommendation surfaces personalized titles based on long-term taste profiles and session behavior. In a streaming context, however, a member issuing a search often has partially formed intent, and purely query-relevant results may miss opportunities for discovery. The paper explores how Netflix's search experience can be enriched by injecting recommendations that adapt in real time to the in-session context — the query, the results shown, and the member's interactions during the session.
Core Contributions
Method / System Architecture
The system follows a high-level pipeline typical of industrial retrieval stacks:
1. Input and representation: the issued query, the member's profile, and in-session interaction context (e.g., rows browsed, titles clicked) are encoded as signals. 2. Candidate generation: two pools are produced — query-matched search results and personalized recommendation candidates conditioned on the session context. 3. Ranking and merging: candidates from both sources are scored and merged into a single results page presentation. 4. Inference strategy: session-adaptation updates recommendation relevance as the session progresses, enabling iterative refinement rather than one-shot retrieval.
Evaluation
> Specific numerical results are not reproduced here; consult the original PDF (DOI: 10.1145/3523227.3547407) before citing quantitative figures.
Key Takeaways
1. Architecture: a cascade of candidate generation, ranking, and merging remains the dominant pattern; the novelty lies in treating retrieval-source mixing as a session-aware decision. 2. Data: session logs are as valuable as long-term taste profiles for interpreting ambiguous search intent. 3. Evaluation: offline relevance metrics alone do not capture in-session discovery benefits — online experimentation is essential. 4. Product: latency, catalog coverage, and explainability constrain how aggressively recommendations can replace query-exact results.
Limitations and Future Work
As summarized from public metadata, potential limitations include dependence on Netflix-proprietary session logs (limiting external reproducibility), the challenge of balancing query relevance against recommendation serendipity, and generalization to catalogs with different content structures. Future directions include deeper joint modeling of query understanding and session-level ranking, and multi-objective optimization balancing search precision with exploration.
Cross-references
Glossary
| Term | Meaning | |------|---------| | IR | Information Retrieval | | In-session adaptation | Adjusting recommendations using signals generated within the current browsing session | | Candidate generation | Producing a manageable pool of items for ranking | | Online A/B test | Randomized controlled experiment measuring real user impact | | nDCG | Normalized Discounted Cumulative Gain, a ranking-quality metric |