The Contest Behind the Feed: Optimal Contest for Recommender Systems

Information Systems and Operations Management

Intervenante: Negin Golrezaei (MIT Sloan)

Salle T004

Abstract

We study a contest-design problem for an online platform’s recommender system that wants to incentivize costly quality-improving effort from multiple content producers. We focus on rank-based exposure rules that recommend content according to its quality rank: the platform chooses how much recommendation probability (exposure) to allocate to each rank. The platform’s goal is to select an exposure rule that maximizes a weighted combination of user welfare and platform quality (with a weight between 0 and 1), evaluated at the equilibrium induced among producers.

Our main result shows that the optimal exposure rule has a strikingly simple and robust structure with three exposure levels: it gives a potentially high exposure to the top-ranked content, zero exposure to the last-ranked content, and the same exposure to every intermediate rank. In other words, equal treatment among the middle ranks—and hence a form of fairness in exposure—emerges endogenously from optimal design.

This structural characterization extends well beyond the baseline objective. It continues to hold when the platform’s objective is the expected value of a broad family of functions that includes any positive-coefficient polynomial combination of content qualities, capturing standard performance measures such as social welfare, order statistics, and several nonlinear transformations (including inverse S-shaped utilities inspired by prospect theory), which are typically avoided due to analytical difficulty.

In special cases, we identify a sharp phase transition: depending on how quickly effort costs accelerate, the optimal rule collapses to either (i) a winner-take-all policy that always recommends the highest-quality content (“HardMax”), or (ii) a nearly uniform policy that recommends all but the last-ranked content with equal probability (“UniformButLast”). Importantly, we also show that HardMax can perform arbitrarily poorly relative to the optimal rule when effort costs are sufficiently convex, underscoring the value of carefully tailored exposure design.