Integrating Predictive Models into Two-Sided Recommendations: A Matching-Theoretic Approach
How to cite
Otani, Suguru; Noda, Shunya; Sekiya, Kazuki; Komatsu, Yuki; Ohkawa, Sachio. 2026. "Integrating Predictive Models into Two-Sided Recommendations: A Matching-Theoretic Approach." AEA RCT Registry. AEARCTR-0015446. https://doi.org/10.1257/rct.15446-1.2
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
Two-sided platforms require two-sided recommendations to ensure both sides benefit from mutual interest. While conventional one-sided recommenders focus on individual preferences, they ignore how over-recommending attractive users can exacerbate inequality. We propose a data-driven framework for efficient and equitable two-sided recommendations by integrating machine learning-based predictive models (e.g., like probabilities and login rates) with insights from matching theory. Through a simulation study using real-world data from CoupLink, a Japanese online dating service, we demonstrate three main findings: (i) when integrating predictive models using deferred acceptance, employing the probability of successful matches as users' preferences leads to better outcomes, (ii) exposure-constrained deferred acceptance (ECDA), which incorporates an improved capacity design, further enhances equity, and (iii) adopting linear programming-based approaches improves both efficiency and equity beyond what ECDA achieves.
Intervention
The intervention involves implementing a new recommendation algorithm, the Exposure-Constrained Deferred Acceptance (ECDA) algorithm, into a Japanese online dating platform. This algorithm is designed to optimize matching efficiency while balancing equity in match allocations. The intervention is introduced in treatment markets, where the platform’s recommendation system is modified to use ECDA, while the control markets continue using the existing recommendation algorithm.
Experimental design
This study is a two-arm cluster-randomized controlled trial with geographic areas as clusters.
Prior to rollout, one of two areas (Kanto vs. Kansai–Tokai) was randomly assigned to receive the ECDA recommendation algorithm, while the other continued using the existing algorithm.
The primary estimand is the intention-to-treat (ITT) effect of ECDA on pre-specified marketplace outcomes during the intervention window.
The experiment runs for two weeks and we track post-treatment activity to obtain post-engagement data. The primary outcome measures include:
- Total Number of Dates (efficiency metric)
- Unique Dating Users (equity metric)
- Proposer and Receiver Gini Index (inequality metric)
- User Engagement Activity (retention measure)
- Total Number of Effective Users (post engagement effectiveness)
The estimation identifies the impact of the ECDA algorithm by comparing pre- and post-intervention differences between treatment and control groups, with market-specific fixed effects controlling for baseline differences.
Design details
Cited by
No citation data available for this registration yet.