Deep Neural Networks for YouTube Recommendations
Covington, Adams and Sargin (2016) describe YouTube's two-stage recommender, whose ranking stage predicts expected watch time. The paper is a description of the 2016 system hosted by Google Research, not an independent audit.
For a psychology course
Separate the documented system or audit from an inference about human behavior. Identify the objective, observation and comparison in this record, then ask what additional evidence would be needed for a causal claim about people.
This record was transferred from Popular Psychology with its original claims, sources and reading limits. Its historical review date does not establish that the current platform works the same way.
- Case type
- Platform
- Context
- Consumer
- Organizations
- Youtube, Google
What happened
In 2016 Paul Covington, Jay Adams and Emre Sargin presented a paper at the ACM Conference on Recommender Systems describing YouTube's recommendation system. It is hosted on Google Research's site. Its abstract calls the system among the largest and most sophisticated in existence and says the paper shares lessons from building and maintaining it. It is a company description of its own 2016 system.
The psychological idea
The system has two stages. A candidate generation network uses a user's watch history to narrow millions of videos to a few hundred. A ranking network then scores those candidates. According to two summaries of the paper, the ranking stage is trained with weighted logistic regression to predict expected watch time rather than clicks.
In practice
Two summaries, by The Morning Paper and InfoQ, report the stated reason for the watch-time target: ranking by clicks can promote clickbait or deceptive videos that users do not finish, while watch time better captures engagement. The Morning Paper reports about one billion model parameters and hundreds of billions of training examples. InfoQ reports that final decisions come from live A/B testing. Later versions of the system are not covered.
Why it mattered
The paper is a dated statement of what YouTube's engineers said the ranker optimized in 2016. It lets later claims about the algorithm be checked against a document, though it says nothing about outcomes for viewers.
Limits and qualifications
This is company self-report by the system's own engineers, not an independent audit. The paper's own text could not be read, so details rest on its abstract and two secondary summaries. The summaries give the reason for watch time as avoiding clickbait. They report nothing showing the design aimed to maximise time spent for its own sake or to push viewers toward extreme content, and effects on viewers are outside the paper.
Evidence behind the claims
The paper by Covington, Adams and Sargin appeared in the Proceedings of the 10th ACM Conference on Recommender Systems (2016) and describes YouTube as one of the largest and most sophisticated industrial recommendation systems, using a deep candidate generation model followed by a deep ranking model.
Candidate generation narrows a corpus of millions of videos to hundreds of candidates, which the ranking stage scores; both summaries agree.
The ranking model uses weighted logistic regression to predict expected watch time rather than click probability; both summaries agree.
The summaries report the stated reason: click-based ranking can promote clickbait or deceptive videos that users do not finish, and watch time better captures engagement.
The models learn about one billion parameters and train on hundreds of billions of examples, according to Colyer's summary of the paper.
InfoQ reports that offline metrics guide development but the final decision comes from live A/B testing, which is also used continuously to improve the ranking model.
People involved
- Paul Covington. Author; paper listed on Google Research's site
- Jay Adams. Author; paper listed on Google Research's site
- Emre Sargin. Author; paper listed on Google Research's site
Course connections
- OpenStax Psychology 2e, section 13.4: Human Factors Psychology and Workplace DesignContext. The section discusses this topic.
The section covers human-machine interface design in general; it does not address recommender systems, and the paper names no psychological theory.
Course relevance is not evidence for the historical or corporate claims.
Sources and inspection
Inspection statements below preserve the original reading scope. A source marked as not verified supports no claim by itself.
- [2]Colyer, A. (2016, September 19). Adriancolyer. The Morning Paper. https://blog.acolyer.org/2016/09/19/deep-neural-networks-for-youtube-recommendations/Colyer, A. The Morning Paper. 2016. News. Secondary.
Relevant text checked. Read the summary as returned by the fetch tool: two stages, candidates in the hundreds, watch-time objective, clickbait reason, scale figures. It does not mention A/B testing.
- [1]Covington, P., Adams, J., & Sargin, E. (2016). Deep neural networks for YouTube recommendations. Google Research. https://research.google/pubs/deep-neural-networks-for-youtube-recommendations/Covington, P; Adams, J; Sargin, E. Google Research; Proceedings of the 10th ACM Conference on Recommender Systems. 2016. Journal article. Primary.
Abstract or publisher record only. Read the abstract and venue on the Google Research page, which lists the publication as to appear. The PDF on Google's static host returned compressed binary that could not be read.
- [3]Giamas, A. (2016, September 23). How YouTube’s recommendation algorithm works. InfoQ. https://www.infoq.com/news/2016/09/How-YouTube-Recommendation-WorksInfoQ. InfoQ. 2016. News. Secondary.
Relevant text checked. Read the news summary as returned by the fetch tool: two stages, watch time, clickbait reason, A/B testing. It gives no scale figures.
Access note: The paper's PDF could not be read, so the stated reason for watch time rests on two secondary summaries that agree with each other. Check the exact wording before quoting.
Original independent review record
Claude review agent, independent of the research agent. 2026-10-21. Verdict: pass-with-edits.
- Attributed the watch-time and logistic regression detail to the summaries, since the paper's text could not be opened.
- Moved the A/B testing statement to InfoQ alone and added a claim for it; the Morning Paper summary does not mention it.
- Added a claim on the abstract's description of the system and its scale.
- Replaced course section 6.3 with 13.4 because the paper names no psychological theory and 13.4 could be confirmed; rewrote first-person phrases impersonally.