RTs != Endorsements: Rethinking Exposure Fairness on Social Media Platforms

TL;DR

The paper tackles exposure fairness on social media by arguing that the social environment and user connections fundamentally shape what users see and perceive. It develops a domain-specific exposure-bias lens across potential, activated, and feed-exposed networks, incorporating socio-cognitive biases that color perception. It contributes a framework for measuring exposure bias with multiple metrics, discusses interpersonal dynamics that affect utility, and offers a toy-example mitigation that emphasizes controlling the degree-attribute correlation $\rho_{kx}$ to reduce perceptual distortion. The work highlights practical implications for platform design and policy, advocating context-aware, perception-aligned fairness in personalized timelines.

Abstract

Recommender systems underpin many of the personalized services in the online information & social media ecosystem. However, the assumptions in the research on content recommendations in domains like search, video, and music are often applied wholesale to domains that require a better understanding of why and how users interact with the systems. In this position paper we focus on social media and argue that personalized timelines have an added layer of complexity that is derived from the social nature of the platform itself. In particular, definitions of exposure fairness should be expanded to consider the social environment each user is situated in: how often a user is exposed to others is as important as who they get exposed to.

Figures (1)

• Figure 1: Simple exposure example. On the left the user is exposed to two users with a globally minority trait and one with the majority trait. On the right the same user, under a different feed is exposed to three users with the minority trait.