How Streaming Algorithms Decide What You Watch: 20 Methods Explained
Streaming services such as Netflix, Hulu, Disney+, Amazon Prime Video, Max, Spotify, and YouTube utilize algorithms for their recommendation systems, with Netflix reporting that these algorithms account for approximately 80% of their viewership. The development of these techniques traces back to the Netflix Prize (2006-2009) and incorporates methods like collaborative filtering, content-based filtering, matrix factorization, and completion data tracking. Behavioral insights, including pausing and skipping actions, take precedence over user ratings. Recommendations are shaped by factors like the time of day, the type of device, and microgenres. Netflix employs around 2,000 taste clusters and session-based models, while onboarding selections help tackle the cold start problem. YouTube leverages reinforcement learning to enhance long-term engagement, and licensing agreements along with seasonal changes also play a role in recommendations.
Key facts
- Netflix says recommendations drive about 80% of what's watched.
- Netflix Prize ran from 2006 to 2009 with a $1 million reward for improving recommendation accuracy by 10%.
- Collaborative filtering recommends based on similar viewers' patterns, not content attributes.
- Content-based filtering uses metadata tags like genre, cast, and tone.
- Matrix factorization breaks down viewing data into latent factors.
- Completion data is a strong signal; binge patterns are tracked.
- Explicit ratings like thumbs up carry less weight than behavioral data.
- Microgenres are narrow categories like 'understated dramas'.
- Thumbnail personalization tests multiple images per title.
- Netflix maintains about 2,000 taste clusters.
- Session-based models weigh recent viewing heavily.
- Cold start problem affects new accounts and new titles.
- Exploration inserts out-of-pattern recommendations to avoid loops.
- Trending rows reflect aggregate recent activity, often regionally.
- Licensing agreements constrain recommendations by country.
- Seasonal adjustments promote timely content.
- Reinforcement learning optimizes for long-term engagement, used by YouTube.
Entities
Institutions
- Netflix
- Hulu
- Disney+
- Amazon Prime Video
- Max
- Spotify
- YouTube
Sources
- Quartz —