conceptinterest-graph-recs
User-controllable recommender
Notes
User-controllable recommender
One-line summary: Recsys prototype where the user issues explicit controls (or edits a personal graph) so the system can break a filter bubble without waiting on like/dislike feedback.
The insight
Fetched papers argue that engagement-inferred user representations go stale relative to what the user says they want, and that passive feedback is too slow to escape a bubble. UCRS (Wang et al., SIGIR 2022) adds alerts + four control commands + on-the-fly counterfactual inference. PKG adaptation (Spadea & Seneviratne, ISWC 2025 companion) edits a user-side knowledge graph before generation. Neither study uses YouTube’s production feed. From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs.
Evidence
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: UCRS “enables users to actively control the mitigation of filter bubbles” and “can respond to the controls and adjust the recommendations on the fly.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Experiments on DIGIX-Video, Amazon-Book, and ML-1M — “not YouTube’s production feed.”
- From 2026-08-21-autoresearch-user-owned-interest-graph-youtube-recs: Personalized soft PKG adaptation raised Out-PIE recommendations from 0.2517 to 0.3237 on Food.com; prompt-only instructions did worse than no adaptation (Out-PIE 0.1925).
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Wang et al. SIGIR 2022 confirmed via Consensus metadata (92 citations); “Experiments on three datasets validate that the UCI framework can effectively recommend more desired items based on user controls.”
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Mysore et al. (SIGIR 2023, LACE) — a user study reports people “could improve the quality of recommendations they received by interacting with the editable profile.” Text recommendations; not YouTube.
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Harambam et al. (RecSys 2019, N=21 news readers) — “an intelligible user profile … coupled with possibilities to influence the recommendation algorithms is highly valued,” and some mechanisms are met with distrust about the control they actually offer.
- From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs: Ain et al. (LAK 2024, N=31) — student-controlled PKG recs beat a content-based baseline on perceived accuracy, novelty, diversity, satisfaction, and use intention. Education, not YouTube.
- From 2026-08-21-autoresearch-best-architecture-for-a-rec-system-web-app: Knowledge-based recommenders “use structured representations of item features, user requirements, and domain rules … independent of past interaction data”; user models “can also be represented as graphs.” See knowledge-based-recommender.
Contradictions / tensions
- Supports the mechanism (user-side graph/commands can steer recs). Does not license “better than YouTube’s feed.”
- Roth et al. (2020) abstract says the empirical literature generally finds bubbles more with explicit / user-declared recommenders than implicit ones. See explicit-vs-implicit-recommendation. From 2026-08-21-academic-research-user-owned-interest-graphs-vs-platform-recs.
Open questions
Related
Referenced by