medium convictionactive · updated 2026-08-21T00:00:00.000Z
YouTube suggestion graph → confinement
Non-personalized YouTube suggestions induce a latent recommendation graph; that mean-field landscape is often confined, and the most confined graphs sit around high-audience videos. This is topological confinement, not Mozilla regret and not a user-owned-graph test.
The chain
1
Non-personalized YouTube suggestions induce a latent recommendation graph from seed videos.
From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: "crawl YouTube’s non-personalized (“mean-field”) suggestion graph from diverse seeds"
2
That mean-field landscape is often confined in topological, topical, and temporal terms, with the most confined graphs organized around high-audience videos.
From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: "the landscape of what we call mean-field YouTube recommendations is often prone to confinement dynamics"
From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: "with the most confined graphs organized around high-audience videos"
What would falsify this
- Step 2: A later fetched study of non-personalized YouTube suggestion graphs with a disclosed crawl protocol finds no topological or topical confinement from diverse seeds.
Contradictions / tensions
- Different object from youtube-recs-to-regretted-watches (Mozilla user-defined regret on recommended vs searched videos).
- Liu et al. 2025 find limited short-term polarization from rabbit-hole-like perturbations of real YouTube recs — polarization, not topological confinement and not regret. From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs.
Implications
- A YouTube-specific confinement finding exists in the peer-reviewed record, but it is about non-personalized suggestion graphs, not a user-drawn interest graph versus the live homepage.
- Roth’s abstract also points at a literature claim that explicit/user-declared recommenders show bubbles more than implicit/activity ones — see explicit-vs-implicit-recommendation.
Companies
Concepts
Open questions