Academic research: user-owned interest graphs vs platform recs

Generated by /academic-research on 2026-08-21. Synthesized across 3 rounds from 15 peer-reviewed papers (see Provenance). Treat as raw material — review before promoting into a project or thread. Context: vault/threads/interest-graph-recs Abstracts-only limitation (v1): claims below are synthesized from Consensus abstracts + metadata, not full papers. Consensus in this environment returned the top 3 hits per query (20 found per search; only those 3 abstracts were retrieved). Research only. Not a product spec. Follow-up to the 2026-08-21 autoresearch pass on this thread.

Note: YouTube-specific data-access claims (watch-history API closed, Takeout delay, no substitute browse) had thin peer-reviewed coverage in the retrieved abstracts. Those facts stay with the prior official-docs source unless a paper actually states them. No Data Portability API finding is invented here.

Priors

Captured from the open question can-a-user-owned-interest-graph-produce-better-recs and the follow-up brief (Paul via Leo). Not a skip.

• P1. The mechanism (explicit user controls / editable personal knowledge graphs) is supported off-YouTube; the comparison to YouTube’s live homepage / Up Next is untested. Confidence: medium-high.
• P2. Lab UCRS / PKG results are not an answer to the YouTube comparison. Confidence: high (methodological stance).
• P3. Official docs (prior pass) say watch history is not in the Data API, Takeout is a delayed archive, and Developer Policies bar recreating browse. Academic/policy confirmation of those three facts was unknown going in. Confidence: high for the docs; low that peer-reviewed abstracts would restate them.
• P4. Graph-database recommendation approaches exist in the peer-reviewed recsys record; vendor Neo4j docs are not that record. Confidence: medium.

Summary

Peer-reviewed abstracts support user-side control of recommendations — alerts plus commands (Wang et al. 2022), editable concept profiles (Mysore et al. 2023), and student-built personal knowledge graphs (Ain et al. 2024) — in lab, news, text, and education settings. None of the retrieved papers ranks a user-drawn interest graph against the same user’s YouTube homepage or Up Next on an on-intent or regret metric. That comparison remains untested.

Two YouTube-specific papers in this corpus cut a different way. A 2020 crawl of non-personalized YouTube suggestions finds confinement in the recommendation graph and notes that prior work generally locates filter bubbles more in explicit (user-declared) recommenders than in implicit (activity-based) ones (Roth et al. 2020). A 2025 PNAS experiment that perturbs real YouTube recommendations into filter-bubble / rabbit-hole conditions reports limited short-term effects on policy attitudes (Liu et al. 2025). Polarization is not Mozilla-style regret, and neither study tests a user-owned graph.

Graph-based recs are a mature peer-reviewed topic: a 2020 survey groups knowledge-graph recommenders into embedding, connection, and propagation methods (Guo et al. 2020); one 2025 IEEE Access paper implements a Neo4j property-graph, Cypher, Jaccard neighborhood recommender on e-commerce transactions (Oprea et al. 2025). That is graph-DB as a store for mined purchase graphs, not a user-authored YouTube interest graph.

On access, retrieved abstracts discuss Search API instability (Efstratiou 2025) and data-donation / GDPR-copy methods as alternatives to platform APIs (Ohme et al. 2023; Pfiffner et al. 2023; Carrière et al. 2024). No retrieved abstract states that watch history is closed in the YouTube Data API, measures Takeout delay, or analyzes the substitute-browse policy. Those remain official-docs findings from the prior pass. This pass invents no Data Portability API result.

Findings

User-side control works in lab domains; it is not a YouTube-feed result

A 2022 SIGIR paper defines a User-Controllable Recommender System (UCRS): alert the user if they are in a filter bubble, accept four granularities of control command, and revise recommendations at inference time with counterfactual inference that blocks stale user representations. The abstract reports that experiments on three datasets (not named as YouTube’s production feed in this abstract) maintain accuracy while following controls (Wang et al. 2022). This is the same UCRS paper already in the thread’s autoresearch source; the peer-reviewed SIGIR venue is now confirmed via Consensus metadata (92 citations).

A later SIGIR paper, LACE, represents users as a small set of human-readable concepts and lets them edit that profile. Offline evaluation spans three tasks and six datasets (warm-start, cold-start, zero-shot). A user study then reports that people could improve the quality of recommendations they received by interacting with the editable profile (Mysore et al. 2023). That is a user-study quality gain from an editable inferred profile, still not a hand-drawn YouTube graph versus the live feed.

In education, students who marked concepts as “did not understand” and thereby controlled construction of a personal knowledge graph received PKG-based resource recommendations that an online study (N=31) reported as better than a traditional content-based baseline on perceived accuracy, novelty, diversity, usefulness, satisfaction, and use intention. The authors say the degree of control over learner modeling had positive consequences for satisfaction and intention to accept recommendations (Ain et al. 2024). Domain is learning resources, not YouTube.

A 2022 SIGIR demonstration paper constructs a Personal Knowledge Graph from a user’s multi-source mobile data and recommends via a “neural + symbolic” approach across services (Yang et al. 2022). It is a system demo (23 citations), not a YouTube comparison.

Qualitative RecSys work in news (four Dutch focus groups, N=21) finds that an intelligible user profile plus the ability to influence the algorithm is highly valued, especially when control serves personal goals — and that some mechanisms are met with distrust about the control they actually offer (Harambam et al. 2019). That is preference for control, not a measured win over YouTube.

What this does not license. Improving perceived quality or diversity in news, text, education, or unnamed recsys datasets is not a ranking against YouTube’s homepage or Up Next. P2 stands.

YouTube-specific papers do not test a user-owned graph — and one cites a tension

Roth, Mazières, and Menezes crawl YouTube’s non-personalized (“mean-field”) suggestion graph from diverse seeds. They report confinement in topological, topical, and temporal terms, with the most confined graphs organized around high-audience videos. The abstract also states that recent empirical studies “generally suggest that filter bubbles may principally be observed in the case of explicit recommendation (based on user-declared preferences) rather than implicit recommendation (based on user activity)” (Roth et al. 2020). That sentence is a literature claim in their abstract, not a test of a user-drawn interest graph. It is a live tension for this thread: a user-owned graph is an explicit preference object. If the cited literature holds, explicit graphs can confine rather than center. This pass did not retrieve the underlying studies Roth alludes to.

Liu et al. built a YouTube-like interface over real YouTube videos and recommendations and experimentally manipulated the algorithm toward ideologically balanced versus slanted (filter-bubble / rabbit-hole) choices — four experiments, nearly 9,000 participants, over 130,000 manipulated recommendations. They report limited causal effects on policy attitudes from even heavy-handed short-term perturbations, and they argue the burden of proof for algorithm-induced polarization has shifted (Liu et al. 2025). Outcome is political polarization, not Mozilla-style regret or on-intent centeredness. The design perturbs YouTube’s existing recs; it does not replace them with a user-drawn graph.

A 2023 systematic review of filter bubbles in recommender systems reports evidence that bubbles exist, names several contributing biases, and argues that incorporating diversity can help (Areeb et al. 2023). It is a recsys-wide review (89 citations), not a YouTube-feed bake-off.

Graph-database rec approaches in the peer-reviewed record

A highly cited TKDE survey (1,013 citations) of knowledge-graph-based recommenders groups methods into embedding-based, connection-based, and propagation-based, and treats the KG as side information to ease sparsity and cold-start and to support explanation (Guo et al. 2020). That is the peer-reviewed map of how graphs are used in recsys. It is not a ranking of Neo4j versus RDF stores, and the graphs in this literature are typically mined item/entity graphs, not user-authored interest maps.

One 2025 IEEE Access paper does use a property-graph engine for recs: customers and products as nodes, purchases as edges, Jaccard over shared purchases, Cypher queries, Neo4j storage, NetworkX visualization. The authors batch relationships with UNWIND and report interactive performance and “high-quality” recommendations on transaction data, validated by comparing Jaccard scores with the Surprise library (Oprea et al. 2025). Domain is e-commerce. Five citations. This is neighborhood collaborative filtering stored in Neo4j, not a user-drawn YouTube graph.

A 2019 ACM Computing Surveys taxonomy (157 citations) compares 51 graph-database systems (including Neo4j, OrientDB, Virtuoso) across document/tuple/native/object stores and RDF versus labeled property graph models, plus sharding and ACID (Besta et al. 2019). Useful as a peer-reviewed graph-DB map. It is not a recommender-quality paper.

YouTube / platform data access in the academic record — what abstracts actually say

Efstratiou ran identical YouTube Data API Search queries for 12 weeks. The Search endpoint returned highly variable results; the API appears to randomize returned videos by relative topical popularity, making representative historical video samples “nearly impossible,” especially off-peak, and may prioritize shorter, more popular videos (Efstratiou 2025). That is a peer-reviewed measurement of Search API instability, not a statement that watch history is closed.

Ohme et al. compare three digital-trace collection methods for social-media effects research — APIs, data donations, and (screen) tracking — and discuss uncertainties because DTD is not designed for research (Ohme et al. 2023). Pfiffner et al. (national survey, n=833) report higher willingness to donate YouTube data than Facebook, Instagram, or Google, with willingness rising when perceived sensitivity is lower, relevance higher, and request/submit control higher (Pfiffner et al. 2023). Carrière et al. give a GDPR-grounded data-donation workflow: the participant requests a copy from the data controller, then chooses what to share; they warn that poorly executed steps threaten study quality (Carrière et al. 2024).

What this pass does not claim. No retrieved abstract names watchHistoryNotAccessible, measures Takeout minutes-to-days delay, or quotes the Developer Policies substitute-browse rule. Donation papers treat a GDPR “copy of your data” as a research method; they do not evaluate Google Takeout latency or the Data Portability API’s YouTube scopes. Do not read a Data Portability finding into this file.

Priors check

• P1 — Confirmed / refined. Additional peer-reviewed user-control and PKG papers (LACE, learner PKGs) still support the off-YouTube mechanism. Liu 2025 is YouTube-specific but tests polarization from perturbed platform recs, not a user-owned graph versus the live feed. The live-feed comparison remains untested.
• P2 — Confirmed. No retrieved paper answers “better / more centered than YouTube’s feed.” Treating UCRS / PKG / LACE as that answer would be a category error.
• P3 — Untested in this academic corpus. The three official-docs constraints are not restated in the retrieved abstracts. Related academic findings (Search API variability; donation as an API alternative; higher YouTube-donation willingness) do not substitute for them.
• P4 — Confirmed. Guo’s survey plus Oprea’s Neo4j implementation plus Besta’s graph-DB taxonomy are the peer-reviewed graph/rec record this pass actually retrieved.

Surprising finding: Roth’s abstract attributes filter-bubble evidence more to explicit/user-declared recommenders than to implicit/activity ones. That is the opposite of the hope that a user-drawn graph automatically centers the feed. It is a literature pointer, not a YouTube user-graph experiment.

No prior is contradicted in a way that requires /calibrate. P1 is refined (YouTube papers exist; they still don’t answer the question). P3 stays a docs finding.

Contradictions and open questions

• Explicit graphs may confine. Roth’s abstract says the empirical literature generally finds bubbles more with user-declared (explicit) recommenders than with activity-based ones (Roth et al. 2020). UCRS / LACE / learner-PKG abstracts treat explicit control as a fix for over-personalization (Wang et al. 2022; Mysore et al. 2023; Ain et al. 2024). Both can be true if “explicit” means a static declared profile versus an editable control surface. This pass did not retrieve the studies Roth cites, so the tension stays open.
• Polarization ≠ regret ≠ centeredness. Liu et al. find limited short-term polarization from rabbit-hole-like YouTube recs (Liu et al. 2025). That does not adjudicate Mozilla-style regretted watches or on-intent match.
• Still no user-drawn-graph vs live YouTube feed study in the retrieved record. That is the unresolved core of can-a-user-owned-interest-graph-produce-better-recs.
• Watch-history / Takeout / substitute-browse are not restated in these abstracts. Peer-reviewed Search-API and donation papers do not replace the official-docs source.
• Graph-DB recs ≠ user-owned interest graph. Oprea’s Neo4j system is purchase-neighborhood CF (Oprea et al. 2025). Guo’s survey is KG-as-side-information (Guo et al. 2020).
• Consensus retrieval cap. Each query reported 20 hits and returned 3 abstracts. A fuller paid Consensus page might surface a YouTube-specific user-graph comparison this pass did not see. Absence here is absence in the retrieved abstracts, not a proof the paper does not exist.

Provenance

Rounds run: 3 of 3

Sub-questions by round:

Round 1 (broad survey):

1. Do user-owned or user-controlled interest graphs produce better or more centered recommendations than platform engagement-optimized recommenders, including any comparison to YouTube?
2. What knowledge-graph and graph-database recommendation approaches exist in the peer-reviewed recsys literature?
3. What do academic or policy papers report about YouTube watch-history access, data-export delays, and API limits for independent recommendation research?

Round 2 (drill-down):

1. Do explicit user-declared preferences produce more or less filter-bubble confinement than implicit engagement-based recommenders? — targeting Roth’s literature claim
2. Peer-reviewed evaluations of interactive / controllable / editable-profile recommenders on quality, diversity, or regret — targeting the gap beyond UCRS
3. Data donation, Google Takeout, or GDPR portability as methods for obtaining YouTube watch history — targeting the access-gap academic record

Round 3 (resolve remaining uncertainty):

1. Neo4j / property-graph / RDF-store recommendation implementations in the peer-reviewed record — targeting graph-DB (not just KG-embedding) recs
2. Academic statements that YouTube watch history is inaccessible via the official API — targeting P3 restatement
3. User-owned / personal knowledge graph recommendation evaluations — targeting PKG beyond the thread’s existing Food.com paper

Papers reviewed (15 total; R1: 6, R2: 5, R3: 4):

Round 1:

• User-controllable Recommendation Against Filter Bubbles — Wenjie Wang et al. (2022). Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval — 92 citations. UCRS + UCI; three datasets; not YouTube production.
• Short-term exposure to filter-bubble recommendation systems has limited polarization effects: Naturalistic experiments on YouTube — Naijia Liu et al. (2025). Proceedings of the National Academy of Sciences — 16 citations. Perturbs real YouTube recs; polarization, not user-owned graph.
• Tubes and bubbles topological confinement of YouTube recommendations — Camille Roth et al. (2020). PLoS ONE — 40 citations. Mean-field YouTube suggestion graphs confine; cites explicit-vs-implicit bubble literature.
• A Survey on Knowledge Graph-Based Recommender Systems — Qingyu Guo et al. (2020). IEEE Transactions on Knowledge and Data Engineering — 1,013 citations. Embedding / connection / propagation taxonomy.
• On YouTube Search API Use in Research — Alexandros Efstratiou (2025). Proceedings of the 2025 ACM Internet Measurement Conference — 4 citations. Search endpoint unstable; historical samples hard.
• Digital Trace Data Collection for Social Media Effects Research: APIs, Data Donation, and (Screen) Tracking — Jakob Ohme et al. (2023). Communication Methods and Measures — 156 citations. API vs donation vs tracking.

Round 2:

• Editable User Profiles for Controllable Text Recommendations — Sheshera Mysore et al. (2023). Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval — 37 citations. LACE editable concepts; user study quality gain.
• Designing for the better by taking users into account: a qualitative evaluation of user control mechanisms in (news) recommender systems — Jaron Harambam et al. (2019). Proceedings of the 13th ACM Conference on Recommender Systems — 97 citations. News focus groups; control valued, sometimes distrusted.
• Filter bubbles in recommender systems: Fact or fallacy—A systematic review — Q. Areeb et al. (2023). Wiley Interdisciplinary Reviews: Data Mining and Knowledge Discovery — 89 citations. Systematic review; diversity as mitigation.
• Leveraging Data Donations for Communication Research: Exploring Drivers Behind the Willingness to Donate — Nico Pfiffner et al. (2023). Communication Methods and Measures — 34 citations. Higher willingness to donate YouTube than FB/IG/Google (n=833).
• Best practices for studies using digital data donation — Thijs Carrière et al. (2024). Quality & Quantity — 23 citations. GDPR copy-request workflow; no Takeout latency number.

Round 3:

• Graph-Based User-Centric Recommender System Using Neo4j, Cypher, and Jaccard Similarity in the Field of e-Commerce — S. Oprea et al. (2025). IEEE Access — 5 citations. Property-graph CF in Neo4j; e-commerce transactions.
• Demystifying Graph Databases: Analysis and Taxonomy of Data Organization, System Designs, and Graph Queries — Maciej Besta et al. (2019). ACM Computing Surveys — 157 citations. 51-system graph-DB taxonomy; RDF vs labeled property graph.
• Learner Modeling and Recommendation of Learning Resources using Personal Knowledge Graphs — Qurat Ul Ain et al. (2024). Proceedings of the 14th Learning Analytics and Knowledge Conference — 24 citations. Student-controlled PKG; N=31 user study vs content-based baseline.
• PKG: A Personal Knowledge Graph for Recommendation — Yu Yang et al. (2022). Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval — 23 citations. PKG demo; neural+symbolic; not YouTube.

Seen in Consensus top-3 but not retained (near-duplicates or off-spine): Peng et al. 2024 KGCFRec (Electronics); Hui et al. 2021 ReBKC (Applied Intelligence); Sui et al. 2022 YouTube-as-research-tool (Digital Health); Ping Liu et al. 2021 news-rec filter bubbles (The Web Conference); Ain et al. 2025 educational user-control levels (unknown journal); Abhilash et al. 2024 Jena vs Neo4j for interactive narratives (not recs); Castaldo et al. 2024 fake-view corrections (Scientific Reports); Campos-Moinier 2026 YouTube API tutorial (Behavior Research Methods, 0 citations); Yuan et al. 2025 PKGRec federated recs (CIKM).

Tools used: mcp__consensus__search (Consensus — Semantic Scholar, PubMed, Scopus, ArXiv). exclude_preprints: true on every query (user asked for peer-reviewed sources). Filters applied: exclude_preprints: true only. No year / SJR / study-type / domain filters. Generated: 2026-08-21 13:20 UTC