YouTube recommendation system
YouTube recommendation system
One-line summary: Two official surfaces (homepage, Up Next) trained on tens of billions of engagement and survey signals; YouTube and mozilla-foundation disagree on whether those recs stay on-intent.
The insight
YouTube describes recommendations as predicting videos the viewer wants, not as a social graph and not as a fixed recipe. Homepage ≈ watch history; Up Next ≈ current video. The 2016 Covington/Adams abstract is the two-stage (candidate generation, then rank) architecture still cited. From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs.
Evidence
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Recommendations “drive a significant amount of the overall viewership on YouTube, even more than channel subscriptions or search.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Signals include “clicks, watchtime, survey responses (‘valued watchtime’ = 4–5 star survey ratings), sharing, likes, and dislikes” and “over 80 billion pieces of information.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Adding watchtime in 2012 produced “an immediate 20% drop in views.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: “Up Next uses the video currently playing as the main signal; homepage primarily uses watch history.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: YouTube reports a 70% drop in U.S. watchtime on non-subscribed recommended borderline content after 2019 demotion, and recommended borderline consumption “significantly below 1%.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Mozilla — “71% of regret reports were recommended videos, and recommended videos were 40% more likely to be regretted than searched-for videos.”
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Liu et al. (PNAS 2025) — four experiments, nearly 9,000 participants, over 130,000 manipulated real YouTube recommendations; “limited causal effects on policy attitudes from even heavy-handed short-term perturbations.” Polarization, not regret.
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Roth et al. (PLoS ONE 2020) — non-personalized (“mean-field”) YouTube suggestion graphs “often prone to confinement dynamics,” most confined around high-audience videos. See youtube-suggestion-graph-to-confinement.
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Areeb et al. (2023 systematic review) — “evidence of filter bubbles in RSs”; diversity as a proposed mitigation. Recsys-wide, not a YouTube-feed bake-off.
The chain
Recommendations (not search or subscriptions) drive a large share of viewing. Homepage is watch-history-based; Up Next is current-video-based. Mozilla's regretted watches land mostly on recommended videos — a different metric than YouTube's demoted-borderline watchtime.
Canonical: youtube-recs-to-regretted-watches.
Contradictions / tensions
- YouTube: independent researchers (unnamed in the fetched post) conclude recs do not steer toward extreme content; borderline recommended watchtime is a small, falling share.
- Mozilla: most regretted videos arrived via recommendations. Different metric than “borderline.”
- Liu 2025: short-term rabbit-hole-like YouTube recs have limited effects on policy attitudes. Polarization ≠ Mozilla regret ≠ on-intent centeredness. From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs.