Adaptive Recommendation Quotas Adaptive Recommendation Quotas for Gig Economy Platforms
How to cite
Otani, Suguru; Noda, Shunya; Sekiya, Kazuki; Komatsu, Yuki; Fujii, Yuki; Ozeki, Shunsuke. 2026. "Adaptive Recommendation Quotas Adaptive Recommendation Quotas for Gig Economy Platforms." AEA RCT Registry. AEARCTR-0017620. https://doi.org/10.1257/rct.17620-1.0
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
This paper studies recommender systems as mechanisms for allocating exposure of items to users, and asks how we can design exposure allocation to achieve more efficient and more equitable matching outcomes. We analyze this problem in the context of Timee, Japan's largest spot-work platform, where workers favorite job templates to receive notifications when offerings are posted. In this environment, we show that conventional recommender systems optimized for conversion metrics such as clicks or favoriting probabilities generate misdirected concentration: they over-recommend job templates that are popular among workers but have limited labor demand and may not lead to employment opportunities. To address this problem, we propose quota-based exposure control mechanisms, including an adaptive quota rule that responds to posting activity and unfilled capacity, as well as a fully parallelizable \emph{Thresholded Exposure Control} (TEC) algorithm suitable for large-scale deployment. In simulations calibrated to 2024 Hokkaido data, job fulfillment rates increase from 57.0\% under a greedy recommender to 63.7\% with static quotas and to 70.6\% with TEC.
Intervention
We experimentally evaluate a recommender algorithm that adaptively allocates exposure across job templates on Japan’s largest spot-work platform, Timee. The treatment prefecture (Aomori) received the proposed Thresholded Exposure Control (TEC) recommender as default in the “template recommendation” tab, while the control prefecture (Iwate) continued using the baseline Greedy recommender optimized for favoriting probabilities. The intervention ran from January 12 to February 11, 2026.
Experimental design
This is a cluster-randomized controlled trial with prefectures as clusters. Prior to deployment, we randomly assigned Aomori and Iwate 1:1 to treatment (TEC) and control (Greedy) using a reproducible script with a fixed seed; the assignment log (timestamp, seed, and code) is archived. We run a prefecture-level, two-arm cluster RCT on Timee to evaluate an exposure-aware recommender. Aomori is assigned to treatment (TEC recommender as default in the template-recommendation tab); Iwate serves as control (status-quo greedy recommender). Outcomes are observed daily at the prefecture level and mirror the simulation targets: (i) average recommended workers per template, (ii) subscribers per template, and (iii) fulfillment rate (confirmed offerings / posted capacity). Effects are estimated via difference-in-differences with prefecture and date fixed effects; the prefecture-level rollout is chosen to mitigate interference in a shared marketplace. Also, we examine the effects on impression-level choice.
Design details
Cited by
No citation data available for this registration yet.