Instead of YouTube’s feed
User-side controls steer recommenders in other catalogs. Nobody has run a user-drawn graph against YouTube’s homepage. One YouTube paper even points the other way: declared preferences may confine more than watch history does.
YouTube says recommendation — not search, not subscriptions — is what people watch most. A person who wants a healthier loop has a plausible alternative: keep a graph of the people, topics, and works they actually care about, and recommend from that instead of from watch history. The question is whether that alternative is supported. It is not, yet, on YouTube itself.
The feed, not search
Cristos Goodrow, in a 2021 YouTube engineering post, described two ranking problems. The homepage ranks from watch history. The “Up Next” panel ranks from the video already playing. Both, he wrote, use hundreds of signals; the company cited more than 80 billion pieces of information a day. Recommendations drive more viewership than search or subscriptions. The same post said YouTube does not recommend videos to steer people toward extreme political content. After changes in 2019, YouTube reported that U.S. watchtime on non-subscribed recommended “borderline” content fell by 70 percent, and that recommended borderline consumption sat significantly below 1 percent.
That is YouTube describing YouTube. It is a first-party account of architecture and of one internal metric. It is not an evaluation of whether a user-built graph would rank better for a person who already knows what they want.
Regret is not borderline, and polarization is not regret
Mozilla’s YouTube Regrets project measured something else. A browser extension, RegretsReporter, collected user-flagged regrets from July 2020 through May 2021: 37,380 volunteers, 3,362 reports from 1,662 of them, across 91 countries. Seventy-one percent of the videos people marked as regretted had reached them through recommendations. Recommended videos were about 40 percent more likely to be regretted than searched videos. In countries where English is not the primary language, the regret rate was 60 percent higher — 17.5 versus 11.0 per 10,000 videos.
The sample is volunteer, not a random panel. Mozilla is measuring user-labeled regret, including off-topic and harmful recs — closer to a rabbit hole than to YouTube’s classifier-demoted “borderline” bucket. Those are different objects. Neither source refutes the other on the other’s terms.
A 2025 PNAS paper then ran a third test. Naijia Liu and colleagues built a YouTube-like interface over real YouTube videos and recommendations, and they pushed the algorithm toward filter-bubble and rabbit-hole conditions — four experiments, nearly 9,000 people, more than 130,000 manipulated recommendations. Even heavy-handed short-term perturbations, they report, had limited effects on policy attitudes. Polarization is not Mozilla’s regret, and it is not “on-intent.” The design also leaves YouTube’s ranker in place. It does not replace the feed with a graph the user drew.
Steering a ranker you own
A personal interest graph is a local store of entities the user names — people, topics, works — and typed links among them. Recommendations then hang off those edges, not off a platform log. The peer-reviewed record supports user-side control in other catalogs. Wang and colleagues (SIGIR 2022) built UCRS: alerts, four control commands, and on-the-fly counterfactual inference, on three datasets that are not YouTube’s production feed. Mysore and colleagues (SIGIR 2023) let people edit a short list of human-readable concepts; in a user study, those edits improved the recommendations they received. In a learning-analytics study (Ain et al., 2024, 31 students), people who marked concepts they did not understand — and thereby built a personal knowledge graph — rated the resulting resource recommendations higher than a content-based baseline on accuracy, novelty, diversity, and satisfaction.
Spadea and Seneviratne treated a personalized knowledge graph as an editable preference store and raised Out-PIE recommendations from 0.2517 to 0.3237 on Food.com recipes; prompt-only instructions did worse than no adaptation. Dutch news readers in a 2019 RecSys focus-group paper valued an intelligible profile and a way to push the algorithm — and some of them distrusted whether the knobs were real.
All of that is evidence that a user can steer a ranker when they own the profile. None of it ran against YouTube’s live homepage or Up Next. Transfer to that feed is unverified. A graph inferred from the same watch history YouTube already uses can become another inferred profile — the thing the papers were trying to escape.
Drawing the map can also close it
There is a YouTube paper that cuts against the hope that a user-drawn graph automatically centers the feed. Roth, Mazières, and Menezes crawled YouTube’s non-personalized suggestions from diverse seeds and watched the graph those suggestions induce. The landscape, they wrote, is often confined — topologically, topically, temporally — and the tightest clusters sit around videos that already have the largest audience.
Their abstract also reports a literature claim: filter bubbles show up more with explicit recommendation, based on preferences the user declared, than with implicit recommendation based on what they actually did. A hand-built interest graph is an explicit object. The papers that treat editable profiles as a fix and the papers that treat declared profiles as a bubble can both be right if “explicit” means a static list rather than a knob you can turn. The wiki leaves that open. Roth’s own crawl is not a bake-off against the homepage.
Two kinds of graph, and a third use
The storage question is older than the YouTube one. Knowledge graphs are nodes plus typed edges. Two families dominate: RDF graphs, queried with SPARQL and constrained with OWL or SHACL, and property graphs, queried with Cypher-class languages. A 2019 ACM Computing Surveys paper compared 51 graph-database systems — Neo4j, OrientDB, Virtuoso among them — across those models. Neo4j’s own documentation still splits the product the way a personal store would care about. Community Edition is fully functional — ACID transactions, Cypher, licensed GPLv3 — and the docs call it ideal for learning, do-it-yourself projects, and small workgroups. Enterprise adds clustering, online backup, role-based access, and multiple user databases: production features a single-user graph does not need in order to exist.
What the recsys literature usually means by “graph,” though, is not a notebook the user keeps. A 2020 IEEE TKDE survey of knowledge-graph recommenders groups the methods into embedding, connection, and propagation, and treats the graph as side information for sparsity and explanation. A 2025 IEEE Access paper puts that into Neo4j: customers and products as nodes, purchases as edges, Jaccard over shared buys, Cypher for the walk. That is neighborhood collaborative filtering stored in a property graph. It is not a YouTube interest map someone drew by hand.
The door that will not open
YouTube’s Data API documents a hard wall: watchHistoryNotAccessible. Watch history cannot be retrieved through the API. Default allocation is 100 search.list calls a day — one unit per call — plus 100 videos.insert calls and 10,000 units for everything else. A 2025 Internet Measurement Conference paper ran the same Search queries for twelve weeks and found the endpoint highly variable, apparently randomized by how hot the topic was that week. Representative historical samples were, the author wrote, nearly impossible. The Developer Policies say API clients must not use the services as a substitute for, or a substantially similar service to, any YouTube application, and must not recreate core experiences such as browse without “significant independent value.”
Google Takeout can export YouTube-related archives, including watch history. The export is delayed, the download link expires in about seven days, each archive can be downloaded five times, and time-range export is not supported. It is not a live recommendation API. Communication-methods papers describe a parallel research path — data donation, a GDPR copy the person requests and then chooses to share — and one survey found people more willing to donate YouTube data than Facebook, Instagram, or Google. Those abstracts do not measure Takeout’s delay, and they do not describe a Data Portability API. A Help page mentions that API for third-party apps; its YouTube-history scopes were not fetched.
A user-owned graph can sit beside a dump. It cannot, on the public surface described here, subscribe to the same firehose the homepage uses. For a channel the user already named, the official cheap path is not a search loop. It is channels.list, then the uploads playlist, then playlistItems.list at one unit a page. search.list stays a 100-call daily bucket. Private user or channel data needs OAuth; public metadata can use an API key. Those are a different auth plane from who owns the graph.
Where the map would even live
How you would host the map is a second question, and the fetched sources refuse to crown a stack. A user-drawn graph sits closer to a knowledge-based recommender — explicit constraints, a graph-shaped user model, no waiting on a watch log — than to the mined item graphs the recsys literature usually means. Ownership of that map and placement of the ranker are different axes. Ink & Switch treat the device copy as primary and the server as a relay; sync can ride CRDTs over any channel, including a USB stick. Solid puts the user in a pod and lets apps request typed access, so the application is not the store. A hosted personal-knowledge base keeps the server as store of record. Ranking can sit in the cloud, on the device, or split — RecGPT-Mobile, on Taobao, keeps intent on-device and fetches items remotely. None of those pages is a YouTube bake-off.
Serving is still retrieve, then rank. KGERA adds a roughly 10-millisecond knowledge-graph reason at test time so the graph can change without retraining — on a mined MovieLens item graph, not a map someone drew. K-RagRec retrieves item-graph subgraphs, re-ranks them, and feeds an LLM. Recreating YouTube browse without “significant independent value” is still forbidden. The independent value, if the app exists, is the graph and the ranker — not a clone of homepage or Up Next.
No fetched page designs, measures, or crowns a single best stack for a YouTube-first personal recs app. Local-first, Solid, and a hosted store are cited as different answers, not a ranked menu. The quality question — whether the graph beats the homepage — stays open on its own page.