Bridging in Social Media Feeds Censors Controversial Topics

Smitha Milli, Luke Thorburn, Paul Bouchaud, Yixin Wang, Nikhil Garg, Emma Pierson

Keywords: recommender systems, bridging systems, social media

Abstract: Under many implementations of bridging-based ranking --- an approach to conflict-sensitive ranking that has been used on social media and for AI alignment --- items are ranked highly if they are valued by people who would normally disagree. Here, we explore the effects of bridging-based ranking in an extensive dataset of 70 million URLs shared on Facebook, and show that while a naive implementation of bridging can lead to higher domain quality among the top-ranked URLs, it can also downrank URLs that relate to controversial topics. These findings suggest that when ranking across many topics, bridging may need to be implemented in a topic-aware way if preserving representation across topics is desired.

Submission Number: 19