User-owned interest graph

One-line summary: A graph of entities the user connects on purpose, used as the gate for what media is allowed to surface — the object this thread is researching, not building.

The insight

A knowledge graph stores entities and typed relations so they can be queried and (in the RDF family) reasoned over. A user-owned interest graph is that object with authorship inverted: the user draws the edges, and recommendations only hang off those edges. Grokipedia already names recommendation as a knowledge-graph application; the fetched recsys papers treat an editable user-side graph as a control surface. From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs.

Evidence

• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: “A knowledge graph represents entities as nodes and typed relationships as directed edges… a knowledge graph adds semantics… so the graph can be queried and reasoned over.”
• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Neo4j’s property-graph shape “matches an interest graph one would draw: Person/Topic/Creator/Video nodes and CARES_ABOUT / RELATED_TO / WATCHED edges.”
• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Spadea and Seneviratne treat a Personalized Knowledge Graph as “a structured, editable representation of preferences.”
• From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: “if the graph is inferred from the same watch history YouTube already uses, it can become another PIE.”
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Ain et al. (LAK 2024) — students who controlled PKG construction got resource recs an N=31 study reported better than content-based on perceived accuracy, novelty, diversity, and satisfaction. Domain is learning resources, not YouTube.
• From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Yang et al. (SIGIR 2022 demo) construct a Personal Knowledge Graph from multi-source mobile data and recommend via a “neural + symbolic” approach. System demo, not a YouTube comparison.
• From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: A user-drawn graph is closer to knowledge-based-recommender (explicit constraints / graph-shaped user models) than to mined knowledge-graph-recommender embeddings. Hosting options cited: local-first-software, solid-project pods, or a hosted PKB — not ranked.

The chain

A user-declared graph can steer recs off an engagement-inferred loop in lab settings (UCRS / PKG papers). Whether that beats YouTube's live feed is untested. The existing YouTube loop is watch-history/current-video signals → recommended views → higher Mozilla regret rate than search.

Canonical: youtube-recs-to-regretted-watches.

Contradictions / tensions

• A mined PKG and a hand-drawn graph are different objects. The fetched papers test the first (or explicit commands), not a user-built YouTube interest graph.
• Roth et al. (2020) abstract: recent studies generally locate filter bubbles more in explicit (user-declared) recommenders than implicit (activity-based) ones. A user-owned graph is an explicit object. See explicit-vs-implicit-recommendation. From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs.

Open questions

• can-a-user-owned-interest-graph-produce-better-recs
• what-web-app-architecture-fits-a-user-owned-interest-graph-recommender

Related

• property-graph-vs-rdf
• user-controllable-recommender
• youtube-recommendation-system
• neo4j
• youtube-recs-to-regretted-watches
• knowledge-graph-recommender
• explicit-vs-implicit-recommendation
• local-first-software
• knowledge-based-recommender
• user-owned-graph-to-quota-aware-youtube-serving