Capping how often popular users are recommended improves fair matching: an exposure-constrained deferred acceptance algorithm raised congestion-adjusted matches and receiver-side match rates on a major Japanese dating app while lowering total raw dates, improving early-stage efficiency without hurting downstream engagement.

Main Finding

Integrating ML predictions into two-sided recommendations as a many-to-many matching problem and controlling receiver exposure by expected likes/dates (not just headcount) reduces harmful congestion and improves matching efficiency for most users. The paper (i) introduces a congestion-adjusted outcome metric, “effective dates,” (ii) shows that deferred-acceptance style integrators using dating-rate rankings outperform naïve like-rate ranking, and (iii) proposes Exposure-Constrained Deferred Acceptance (ECDA) — which caps receiver exposure in expected-likes/dates — that increases effective dates and receiver-side dating probability in calibrated simulations and in a large regional field experiment on a major Japanese dating platform.

Key Points

• Problem set-up

  • Two-sided recommendation (proposers → receivers) requires mutual activity and responsiveness. Simple ranking by predicted pairwise match probability concentrates exposure on a small set of receivers, causing congestion and overstated efficiency.
  • Predictions available: login rates λ, proposer→receiver like rates αij, receiver relike rates βij. Dating rate for pair (i,j): δij = λi αij λj βij.
  • Recommendations are encoded by a matrix M with proposer attention caps ci (each proposer sees up to ci receivers).

• New metric: Effective dates

  • Captures congestion: when a receiver accumulates multiple dates, only one is assumed to become an effective relationship.
  • Let µj = Σk δkj Mkj (expected dates for receiver j). Effective dating rate for pair (i,j): δ*ij = ((1 − e−µj) / µj) · δij.
  • Average effective dates discounts matches to overloaded receivers; it highlights the welfare loss from congestion that average raw dates can miss.

• Integrator designs and findings

  • Rank-order input matters: sorting ROLs by dating rate (δij) for DA is better than by like rate.
  • Conventional “One-sided” greedy ranking by dating rate concentrates exposure on top receivers and performs poorly on effective dates.
  • Deferred Acceptance (DA) with appropriate capacity parameters mitigates congestion, but further gains come from ECDA.
  • ECDA defines receiver capacity in terms of expected likes/dates (exposure), not headcount. This directly targets congestion externalities and is computationally tractable (greedy implementations leveraging δij sorting).

• Empirical validation

  • Calibration: simulations use production-grade predictive models from CoupLink (large Japanese dating app, ≈1.5M cumulative users).
  • Simulations: ECDA increases average effective dates and receiver-side dating probability even though it reduces total raw predicted dates relative to One-sided.
  • Field experiment: region-level difference-in-differences (treatment: Kanto; controls: Kansai & Tokai); pre-registered and IRB-approved.
    • ECDA led to a more equitable distribution of likes/dates and increased effective dates versus the incumbent One-sided recommender.
    • Early-stage engagement (likes, upstream activity) increased on both sides.
    • Aggregate downstream outcomes (messaging) were muted or negative when including the extreme right tail of receivers: the top 0.1% of receiver-days had been capturing many low–pass-through likes under One-sided. After excluding that tail, treatment effects are coherent and positive for effective dates, dating probability, and likes for 99.9% of users.

• Practical considerations

  • ECDA is scalable and practical for production deployment.
  • The paper focuses on integrating predictions (assumes reasonably accurate ML estimates) rather than improving the predictions themselves.
  • The “uniform choice” assumption underlying effective dates is a transparency device to discount congestion; it is not presented as a literal behavioral model.

Data & Methods

• Data sources

  • Production prediction outputs (login, like, relike probabilities) and platform activity from CoupLink, a large Japanese dating service.
  • Large-scale regional field deployment with pre-registered evaluation (AEA RCT Registry AEARCTR-0015446) and IRB approval.

• Model and mechanics

  • Recommendation: choose matrix M satisfying proposer capacities ci.
  • Dating probability per pair: δij Mij (with δij = λi αij λj βij).
  • Performance metrics: average dates, average effective dates (congestion-adjusted), dating probability (probability a user gets at least one date).
  • ROL construction and integrators: compare One-sided greedy ranking by δij, DA variants (with headcount capacities), and ECDA (receiver capacity in expected-likes/dates).

• Simulations

  • Calibrated numerical simulations using the platform’s predictive models to evaluate counterfactual integrator performance at scale.

• Field experiment

  • Regional difference-in-differences: Kanto (treatment) vs Kansai & Tokai (controls); outcome windows include recommendation-stage expected outcomes and realized outcomes up to two weeks post-recommendation.
  • Analysis examines heterogeneity and the impact of the extreme high-exposure receiver tail (top 0.1% receiver-days).

• Computation

  • ECDA admits greedy and parallelized implementations leveraging δij sorting; designed for large-scale production use.

Implications for AI Economics

• Integration (not just prediction) matters: How ML outputs are aggregated into platform decisions (integrators/mechanisms) can materially change market outcomes. Platforms need mechanism-design-aware integration of predictive models rather than naive ranking.
• Externalities of concentrated exposure: ML-based rankings that maximize immediate conversion probabilities can impose negative externalities (congestion) on counterparties, reducing realized welfare. Exposure constraints (in expected-likes/dates) internalize such externalities.
• Metrics should be market-aware: Standard one-sided recommender metrics (clicks, predicted matches) can be misleading in two-sided markets. Outcome metrics that incorporate congestion (like effective dates) better capture true value and equity.
• Mechanism design with ML inputs: ECDA exemplifies combining predictive ML with matching-theoretic constraints to create scalable, welfare-improving policies. This mirrors broader AI-economics themes of using predictions as inputs to economic mechanisms rather than as end-products.
• Equity-efficiency tradeoffs: Reducing aggregate raw matches may increase meaningful matches for the many. Policymakers and platforms should weigh aggregate throughput against distributional and downstream quality effects.
• Evaluations need field experiments: Calibrated simulations are useful, but real-world behavioral responses and extreme-tail phenomena (very high-exposure receivers) require randomized or quasi-experimental validation.
• Generalizability and domains: Lessons extend beyond dating to job platforms, freelance marketplaces, and other reciprocal recommendation settings where responsiveness and activity matter and where over-concentration harms market functioning.
• Future directions: modeling strategic behavior (if users adapt to exposure rules), improving behavioural assumptions behind effective-date discounting, jointly optimizing prediction and integrator design, and exploring regulatory implications for platform exposure policies.

Limitations to note - Effective dates use a simplifying uniform-choice assumption to quantify congestion; richer behavioral models may refine the metric. - The analysis assumes available, sufficiently accurate predictions; results depend on prediction quality. - The field evidence is from a dating platform; application to other markets requires domain adaptation and additional validation.

Overall, this paper provides a practical, theoretically grounded framework for turning multi-dimensional behavioral predictions into matching-aware recommendations that control exposure and reduce harmful congestion in two-sided markets.