Deadweight Loss in Recommendation Systems

초록

Content recommender systems can be viewed as platform-mediated allocation mechanisms among users, creators, and the platform. We model platform-side intervention as a wedge that can leave mutually beneficial user-creator interactions unrealized. To operationalize this effect, we construct an empirical Pareto frontier over user-side and creator-side utility and compare the resulting allocations with those obtained when one top-K position is reserved for platform-beneficial exposure. Experiments show that slot reservation can reduce user-side ranking utility and creator-side exposure, while displacing relevant matches retained in the reference allocation. We operationally define deadweight loss (DWL) as the loss incurred when relevant user-creator matches are displaced without preserving either user-side utility or creator-side exposure. © 2026 Copyright held by the owner/author(s).

키워드