Recommendation Feedback Loops Shape Urban Income Segregation and Venue Visitation
Guangyuan Weng, Lin Chen, Ciro Beneduce, Giovanni Mauro, Massimiliano Luca, Takahiro Yabe, Bruno Lepri, Luca Pappalardo
Timely, well-executed simulation study with counterintuitive accuracy–segregation findings and an actionable mitigation, limited by simplified behavior (adoption=1), static venue supply and no real-world validation.
Abstract
Location-based recommender systems shape where people go in cities and learn from the visits they steer, forming a human-AI feedback loop. This feedback loop shapes whom people share places with across income groups and which venues keep attracting customers. Yet platforms judge these systems by their offline accuracy on past visits, a test in which the suggestions never change where people go and which therefore cannot capture long-term effects. Studying these effects in cities is challenging because, unlike online resources, travel constraints and limited venue capacity shape where people can go, while each recommended visit also changes whom other visitors encounter. To capture such features, we build a long-term simulation framework based on fine-grained human mobility data from four U.S. metropolitan areas, with distance and capacity constraints. Focusing on food venues, we study two urban consequences: how unevenly people's exposure is distributed across income groups (experienced income segregation) and the share of venues receiving no visits from simulated residents (venue disappearance). We find that, under every algorithm and in every metropolitan area, at least a fifth of venues stop receiving visits, and that more than half of this loss arises from retraining. More importantly, more accurate algorithms leave people more segregated, and the most accurate algorithm leaves people 15-21% more segregated than they actually are. We then test a re-ranking intervention that favours less popular venues and runs on top of a platform's existing algorithm. It lowers both venue disappearance and segregation across algorithms, largely without directing people to venues where they may feel less comfortable. Our results suggest that impact assessments of recommender systems should account for the feedback loop and report segregation and venue disappearance beyond accuracy.
AI Impact Assessments
(1 model)Impact Assessment: Recommendation Feedback Loops Shape Urban Income Segregation and Venue Visitation
Core Contribution
The paper builds a long-horizon (about three years of weekly retraining) simulation of the human–AI feedback loop for location-based recommenders. It is grounded in Cuebiq GPS traces from four U.S. metros (Boston, San Francisco, Detroit, Chicago) and respects distance and venue-capacity constraints. Non-simulated visitors keep their observed visits, so recommendations to simulated residents also change the income mix other people encounter.
It measures two urban outcomes: experienced income segregation (Moro et al.'s measure) and "venue disappearance," the share of venues with no simulated visits.
The headline findings are:
The conceptual advance is moving recommender evaluation from out-of-loop accuracy to in-loop, distributional, spatially grounded outcomes. The paper adds capacity constraints, social coupling and real GPS schedules to the prior Mauro et al. check-in-based urban simulation.
Methodological Rigor
Strengths
Weaknesses
Potential Impact
The work speaks directly to recommender evaluation (accuracy-only paradigms), the human–AI coevolution literature, urban computing and computational social science. It also speaks to algorithmic auditing under the EU Digital Services Act. The practical message is actionable: report segregation and venue-visibility metrics alongside accuracy. Switching UR-POI to ItemKNN costs about 4–5% NDCG yet lowers both harms, and a simple re-ranker works on top of any base algorithm. Platforms such as Google Maps, Yelp and Foursquare, or auditors and regulators, could in principle use the framework, though it needs proprietary-quality mobility data. Policy relevance is high, but adoption by platforms is uncertain, and the framework's usability depends on access to GPS data like Cuebiq's.
Timeliness & Relevance
It is highly timely. Feedback-loop effects, regulatory systemic-risk assessments, the rise of LLM-based place recommenders, and the experienced-segregation literature all converge here. The authors acknowledge that LLM assistants remain for future work.
Strengths
Limitations
Other Observations
The paper is dense but clearly structured. It reads as an extension of the group's earlier simulation work and sits in a growing cluster of feedback-loop simulation papers. It probably will be cited as a reference for in-loop evaluation of location recommenders and for urban fairness metrics, though not as a foundational primitive.
Generated Nov 2, 2026
Comparison History (0)
Timely, well-executed simulation study with counterintuitive accuracy–segregation findings and an actionable mitigation, limited by simplified behavior (adoption=1), static venue supply and no real-world validation.