[Paper Review] Challenges in Search on Streaming Services: Netflix Case Study

Sudarshan Lamkhede, Sudeep Das|arXiv (Cornell University)|Mar 11, 2019

Mobile Crowdsensing and Crowdsourcing7 references4 citations

TL;DR

This paper investigates the unique challenges of building a search system for streaming platforms like Netflix, emphasizing the integration of recommendations into search to support discovery, handle unavailable content, and enable instant, device-agnostic search across global, multilingual audiences. It proposes a hybrid search-recommendation framework that blends lexical matching with behavioral signals to improve user satisfaction and content discovery beyond traditional IR models.

ABSTRACT

We discuss salient challenges of building a search experience for a streaming media service such as Netflix. We provide an overview of the role of recommendations within the search context to aid content discovery and support searches for unavailable (out-of-catalog) entities. We also stress the importance of keystroke-level instant search experience, and the technical challenges associated with implementing it across different devices and languages for a global audience.

Motivation & Objective

• To address the limitations of traditional information retrieval in streaming media search, where users seek not just specific content but also discovery and serendipitous recommendations.
• To understand and support three distinct user search mindsets: Fetch (specific item), Find (related items), and Explore (broad content areas), each requiring different system behavior.
• To design a search system that seamlessly integrates recommendations with lexical search to handle unavailable content and improve user engagement.
• To enable low-latency, keystroke-level instant search across diverse devices and input methods, especially in multilingual contexts like Hangul input for Korean.
• To ensure high-quality, scalable search and recommendation performance across 22+ localized languages with varying content preferences and query patterns.

Proposed method

• The system combines lexical search (based on title and metadata matching) with collaborative filtering and behavioral signals to generate recommendations for both available and unavailable content.
• It implements a real-time, incremental search interface that updates results after each keystroke, reducing user interaction cost and improving perceived responsiveness.
• The system uses language-specific indexing (e.g., Hangul syllable-level indexing for Korean) to minimize keystrokes and improve input efficiency for non-Latin scripts.
• It applies transfer learning from US English to localize search and recommendation models in new markets, adapting to regional content preferences and query patterns.
• The search pipeline integrates recommendations at multiple levels—within search results, for unavailable entities, and during query understanding—to guide users through content exploration.
• A/B testing and user studies are used to evaluate search quality, interaction efficiency, and user satisfaction across different use-cases and locales.

Experimental results

Research questions

• RQ1 How can search systems on streaming platforms effectively support users who are not searching for a specific title but are instead exploring or discovering content?
• RQ2 What role should recommendations play in search results when the queried content is unavailable for streaming?
• RQ3 How can instant search be implemented efficiently across diverse devices and input methods, especially for languages with complex scripts like Korean?
• RQ4 How can search and recommendation systems be jointly optimized to reduce user interaction cost while increasing content discovery?
• RQ5 What are the key differences in search behavior and system requirements between global streaming platforms and traditional web search?

Key findings

• Over 20% of content discovery on Netflix occurs through search, highlighting its critical role beyond just retrieval.
• The majority of searches (about 75%) are for specific titles (Fetch), but a significant portion involve broader intent (Find and Explore), requiring recommendation integration.
• Users frequently search for unavailable content (e.g., 'Jaws' when looking for shark movies), and the system must surface relevant recommendations to maintain engagement.
• Keystroke-level instant search reduces user effort, but it increases system complexity due to multiple short queries and higher error rates on unreliable networks.
• Language-specific indexing, such as Hangul syllable-level indexing, can reduce the number of required keystrokes by up to 60% compared to character-level input, improving input efficiency.
• Localization of titles (e.g., 'La Casa de Papel' as 'Money Heist') creates mismatches between user queries and indexed content, requiring robust cross-title matching and recommendation strategies.