Abstract

Algorithms personalize social media feeds by ranking posts from the inventory of a user’s network. However, the combination of network structure and user activity can distort the perceived popularity of user traits in the network well before any personalization step. To measure this “exposure bias” and how users might perceive their network when subjected to personalization, we conducted an analysis using archival X (formerly Twitter) data with a fixed inventory. We compare different ways recommender systems rank-order feeds: by recency, by popularity, based one the expected probability of engagement, and random sorting. Our results suggest that users who are subject to simpler algorithmic feeds experience significantly higher exposure bias compared to those with chronologically-sorted, popularity-sorted and deep-learning recommender models. Furthermore, we identify two key factors for bias mitigation: the …

Metadata

publication

International Conference on Advances in Social Networks Analysis and Mining …, 2024

year

2024

publication date

2024/9/2

authors

Nathan Bartley, Keith Burghardt, Kristina Lerman

link

https://link.springer.com/chapter/10.1007/978-3-031-78538-2_3

book

International Conference on Advances in Social Networks Analysis and Mining

pages

38-53

publisher

Springer Nature Switzerland