interest-graph-recs

refreshed 1, created 0, skipped 0

• refreshed interest-graph-youtube-recs ← architecture pass: knowledge-based vs mined KG, local-first/Solid/hosted, playlistItems.list cheap path, retrieve-then-rank. Quality question left open.

Architecture-only ingest from /autoresearch on a rec-system web app over a user-defined interest graph (layers, local-first vs hosted, sync, rec serving, YouTube fetch constraints already known). That source only.

Created entities:

• ink-and-switch (entity): local-first essay; device replica primary.
• solid-project (entity): user-owned pods; apps request typed access.

Created concepts:

• local-first-software (concept): device-primary vs hosted PKB vs Solid.
• on-device-recommender (concept): cloud RS vs ODRS; RecGPT-Mobile split.
• multi-stage-rec-serving (concept): candidate generation → rank → optional test-time KG reason.
• knowledge-based-recommender (concept): explicit constraints / graph user models; distinct from mined knowledge-graph-recommender.

Created mechanism:

• user-owned-graph-to-quota-aware-youtube-serving (3 steps, 6 citations) — user-owned graph primary → cheap playlistItems candidates → rank without cloning browse.
• Updated concepts local-first-software and multi-stage-rec-serving: ## The chain → summary linking user-owned-graph-to-quota-aware-youtube-serving.
• Did not rewrite user-owned-interest-graph ## The chain (existing Canonical remains youtube-recs-to-regretted-watches).

Updated:

• youtube-data-access-constraints: official known-channel playlistItems path; policies re-fetched; OAuth vs API key as a separate auth plane.
• knowledge-graph-recommender: KGERA test-time reasoning; K-RagRec retrieve-then-rerank.
• user-owned-interest-graph: architecture as preference-store hosting; pointer to the new question.
• user-controllable-recommender: knowledge-based explicit / graph user models.
• youtube-platform: official samples + policy re-fetch.
• what-web-app-architecture-fits-a-user-owned-interest-graph-recommender (question, still open): families compared; no single best stack.
• Left can-a-user-owned-interest-graph-produce-better-recs open (rec quality vs feed; not this ingest’s verdict).

Updated index (index): new pages and the architecture source.

Refreshed 1, created 0, skipped 0

• refreshed interest-graph-youtube-recs ← can-a-user-owned-interest-graph-produce-better-recs, user-owned-interest-graph, user-controllable-recommender, youtube-recommendation-system, youtube-recs-to-regretted-watches, youtube-suggestion-graph-to-confinement, youtube-data-access-constraints, property-graph-vs-rdf, knowledge-graph-recommender, data-donation, explicit-vs-implicit-recommendation, youtube-platform, neo4j, mozilla-foundation, google-takeout
• reason: wiki moved after academic ingest (new concepts + Roth mechanism); extend existing topic, no new slug

Created 1, refreshed 0, skipped 0

• created interest-graph-youtube-recs ← can-a-user-owned-interest-graph-produce-better-recs, user-owned-interest-graph, user-controllable-recommender, youtube-recommendation-system, youtube-recs-to-regretted-watches, youtube-data-access-constraints, property-graph-vs-rdf, youtube-platform, neo4j, mozilla-foundation, google-takeout

Academic-research ingest (Consensus abstracts, 15 papers). That source only — not leftover auto queue.

Created concepts:

• knowledge-graph-recommender (concept): Guo 2020 embedding/connection/propagation taxonomy; Oprea 2025 Neo4j instance.
• data-donation (concept): Ohme / Pfiffner / Carrière methods papers; no Takeout-delay or Data Portability finding.
• explicit-vs-implicit-recommendation (concept): Roth 2020 literature claim that bubbles show more with user-declared recs.

Created mechanism:

• youtube-suggestion-graph-to-confinement (2 steps, 3 citations, both partial) — mean-field YouTube suggestion graph → confinement around high-audience videos.
• Did not rewrite youtube-recommendation-system ## The chain (existing Canonical remains youtube-recs-to-regretted-watches; different spine).

Updated:

• user-owned-interest-graph: Ain 2024 / Yang 2022 PKG; Roth explicit-graph tension.
• user-controllable-recommender: Wang venue confirmed; Mysore LACE user study; Harambam 2019; Ain 2024.
• youtube-recommendation-system: Liu 2025 polarization; Roth confinement; Areeb 2023 review.
• youtube-data-access-constraints: Efstratiou 2025 Search API instability; donation papers do not restate watchHistoryNotAccessible.
• property-graph-vs-rdf: Besta 2019 taxonomy; Oprea 2025 Neo4j rec.
• neo4j: Oprea + Besta.
• google-takeout: donation literature is not a latency measurement.
• youtube-platform: Efstratiou + Liu.
• youtube-recs-to-regretted-watches: Liu/Roth as different-metric tensions.
• can-a-user-owned-interest-graph-produce-better-recs (question, still open): more off-YouTube control evidence; live-feed comparison still untested; Roth tension recorded.

Updated index (index): new pages and the academic source.

Seed ingest from /autoresearch on a user-owned interest graph for healthier YouTube recs (graph-DB options, rec quality vs YouTube feed, YouTube data/graph/rec constraints). Wiki had only the open question page before this ingest.

Created entities:

• youtube-platform (entity): homepage + Up Next contrast case; slug avoids YOUTUBE.md basename clash.
• neo4j (entity): Community Edition as the only official-docs engine fetched.
• mozilla-foundation (entity): YouTube Regrets 71% / 40% / 60%.
• google-takeout (entity): batch export, not a live rec API.

Created concepts:

• user-owned-interest-graph (concept): user-drawn graph as rec gate.
• youtube-recommendation-system (concept): official signals; YouTube vs Mozilla tension.
• property-graph-vs-rdf (concept): two KG families; no unofficial bake-off cited.
• youtube-data-access-constraints (concept): watchHistoryNotAccessible, 100 search.list/day, substitute-service policy.
• user-controllable-recommender (concept): UCRS / PKG; not YouTube production.

Updated questions:

• can-a-user-owned-interest-graph-produce-better-recs (question, still open): mechanism supported off-YouTube; vs live feed untested.

Created mechanism:

• youtube-recs-to-regretted-watches (3 steps, 3 citations) — recs drive viewing → history/current-video rankers → Mozilla regret higher on recommended than searched.
• Updated concepts youtube-recommendation-system and user-owned-interest-graph: ## The chain → summary linking youtube-recs-to-regretted-watches

Updated index (index): added all new pages and the source.

Thread interest-graph-recs scaffolded via /thread-create.

• Scope: graph DB options for a personal interest graph + recs; whether a user-owned graph can produce better, more centered recs than YouTube’s feed; YouTube data / graph-shape / rec-approach constraints. Research only — not a product. See SCOPE.md.