Advertising for Demographically Fair Outcomes
Online advertising on platforms such as Google or Facebook has become an indispensable outreach tool, including for applications where it is desirable to engage different demographics in an equitable fashion, such as hiring, housing, civic processes, and public health outreach efforts. Somewhat surprisingly, the existing online advertising ecosystem provides very little support for advertising to (and recruiting) a demographically representative cohort. We study the problem of advertising for demographic representativeness from both an empirical and algorithmic perspective. In essence, we seek fairness in the outcome or conversions generated by the advertising campaigns. We first present detailed empirical findings from real-world experiments for recruiting for civic processes, using which we show that methods using Facebook-inferred features are too inaccurate for achieving equity in outcomes, while targeting via custom audiences based on a list of registered voters segmented on known attributes has much superior accuracy. This motivates us to consider the algorithmic question of optimally segmenting the list of individuals with known attributes into a few custom campaigns and allocating budgets to them so that we cost-effectively achieve outcome parity with the population on the maximum possible number of demographics. Under the assumption that a platform can reasonably enforce proportionality in spend across demographics, we present efficient exact and approximation algorithms for this problem. We present simulation results on our datasets to show the efficacy of these algorithms in achieving demographic parity.