Personalized recommendations in capacity-constrained matching markets can trigger a self-defeating 'recommendation-induced congestion' that erodes the offer-likelihood guarantee they were based on, and does so disparately

matching-marketsrecommender-systemsschool-choicenycalgorithmic-fairnessdeferred-acceptancecongestionarxiv

Using 2024-25 NYC high school admissions data, Chiang, Peng, Lichtenstein, McDaniel, O'Neil, Thomas, Wright, Kleinberg, Tardos & Garg simulate what happens when every eligible student is nudged toward every nearby high-performing program on which they hold a "High" (≥99%) predicted offer likelihood. As adoption of the nudges rises, acceptance rates at the recommended programs among High-likelihood applicants fall sharply — a phenomenon the paper names recommendation-induced congestion: "if they steer too many users toward the same items, then even users who were originally predicted to have a high chance of matching to an item may not, due to increased competition." At 20% nudge adoption, acceptance rates fall below 80% at 7 of 56 programs; at 50% adoption, at 15 of 56.

The effect is not evenly distributed. The authors report that "students with less access to high-performing programs are more likely to be affected by congestion at the programs they are eligible for, due to geographic correlation" — congestion concentrates precisely where applicants already had the fewest nearby options, compounding rather than offsetting existing access disparities.

The underlying mechanism this paper studies — a centralized deferred-acceptance match — is the same stable-matching framework formalized in Gale and Shapley's 1962 algorithm; this paper's object of study is a 2026 recommendation layer sitting on top of that six-decade-old mechanism. See also claim-balanced-capacitated-nudging-guarantees-safe-and-near-optimal-matches for the paper's proposed fix, and entity-nikhil-garg / entity-kenny-peng for the research cluster this sits in.