Primary Submission Category: Randomized Studies
Experimental Design For One-sided Matching Marketplaces
Authors: Nian Si, Chenran Weng, Xiao Lei,
Presenting Author: Nian Si*
One-sided matching markets, prevalent in scenarios where users are matched with other users, are evident in environments like video game platforms and anonymous social networks. Here, participants are matched for interactions such as games or social exchanges. Experimentation (A/B tests) in these markets is challenging due to the interdependence of users’ metrics on their counterparts’ treatment assignments. In this paper, we build a stochastic market model and develop its mean field limit to analyze such experimental dynamics. Our focus is on two randomization strategies: user and match randomization. We demonstrate that, under Markovian conditions and homogeneous users behavior, match randomization provides unbiased estimations but can lead to significant biases when these conditions are not met. Conversely, user randomization shows greater resilience to model inaccuracies. We further propose an associated linear regression estimator that can halve the bias compared to a naive estimator.