Can a user-owned interest graph produce better, more centered recs than YouTube’s feed?

The question

Can a user-owned interest graph produce better, more centered recs than YouTube’s feed?

Why it matters

SCOPE’s rec-quality claim. YouTube’s feed is engagement-inferred (youtube-recommendation-system); the alternative is a graph the user draws (user-owned-interest-graph). “Better / more centered” here means on-intent, fewer rabbit holes — Mozilla-style regret — not only YouTube’s classifier “borderline” bucket.

What we currently believe

The mechanism is supported in other domains: explicit user controls (UCRS, LACE editable profiles) and user-built / editable PKGs can improve perceived quality or reduce over-personalization without a full retrain. The comparison to YouTube’s live homepage / Up Next is still untested. Two YouTube-specific peer-reviewed papers in the academic pass (Liu 2025 polarization; Roth 2020 mean-field confinement) do not run a user-drawn graph against the live feed. Roth’s abstract also points at a literature claim that explicit recommenders may confine more than implicit ones — a live tension, not a resolution. From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs and 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs.

Evidence we have

• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Wang et al. (SIGIR 2022) UCRS — user controls + counterfactual inference; experiments on DIGIX-Video, Amazon-Book, ML-1M, “not YouTube’s production feed.”
• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Spadea & Seneviratne — personalized soft PKG adaptation raised Out-PIE from 0.2517 to 0.3237 on Food.com; prompt-only instructions did worse than no adaptation.
• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: YouTube vs mozilla-foundation do not agree on whether the existing feed stays on-intent (borderline-watchtime vs user-defined regret). See youtube-recs-to-regretted-watches.
• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: even if the graph worked, youtube-data-access-constraints (no watch-history API, 100 search.list/day, substitute-service policy) bound what a third-party loop can retrieve.
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Mysore et al. (SIGIR 2023) — user study: people improved rec quality via an editable concept profile. Not YouTube.
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Ain et al. (LAK 2024, N=31) — student-controlled PKG beat content-based on perceived accuracy, novelty, diversity, satisfaction. Education, not YouTube.
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Liu et al. (PNAS 2025) — naturalistic YouTube rec perturbations; limited short-term effects on policy attitudes. Not a user-owned-graph vs feed ranking.
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Roth et al. (PLoS ONE 2020) — mean-field YouTube suggestion graphs confine; abstract says prior work generally finds bubbles more with explicit/user-declared recs than implicit. See explicit-vs-implicit-recommendation.
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: No retrieved abstract restates watch-history API closure, Takeout delay, or substitute-browse policy, and none compares a user-drawn graph to the live homepage / Up Next.

Evidence we need

• A study (or even a documented personal experiment) that ranks candidates from a user-drawn interest graph against the same user’s YouTube homepage / Up Next on an on-intent or regret metric.
• Fetched Data Portability API scopes, if they change the watch-history path.

How to resolve

Do not treat lab UCRS/PKG/LACE results as an answer. Liu 2025 and Roth 2020 are YouTube-specific but do not rank a user-drawn graph against the live feed. A later pass would still need that comparison; do not invent a Data Portability finding.

Related

• Queue entries: interest-graph-recs (done), interest-graph-recs-academic
• what-web-app-architecture-fits-a-user-owned-interest-graph-recommender (architecture, not rec quality)
• user-owned-interest-graph
• user-controllable-recommender
• youtube-recommendation-system
• youtube-recs-to-regretted-watches
• explicit-vs-implicit-recommendation
• knowledge-graph-recommender
• data-donation