Fairness in social influence maximization

Abstract

Algorithms for social influence maximization have been extensively studied for the purpose of strategically choosing an initial set of individuals in a social network from which information gets propagated. With many applications in advertisement, news spread, vaccination, and online trend-setting, this problem is a central one in understanding how information flows in a network of individuals. As human networks may encode historical biases, algorithms performing on them might capture and reproduce such biases when automating outcomes. In this work, we study the social influence maximization problem for the purpose of designing fair algorithms for diffusion, aiming to understand the effect of communities in the creation of disparate impact among network participants based on demographic attributes (gender, race etc). We propose a set of definitions and models for assessing the fairness-utility tradeoff in designing algorithms that maximize influence through a mathematical model of diffusion and an empirical analysis of a collected dataset from Instagram. Our work shows that being feature-aware can lead to more diverse outcomes in outreach and seed selection, as well as better efficiency, than being feature-blind.