Filters for Social Media Timelines: Models, Biases, Fairness and Implications

Abstract

Social media have a significant impact, on the lifestyle, behaviour and opinion of billions of users. To handle the flow of information between its members, social media developed personalization algorithms that filter the contents that flow into users' timelines. Despite the far-reach of social media filters, such algorithms lack transparency, motivating research to understand and improve its properties. In this thesis, bridging queuing theory, caching models and network utility maximization, we propose a reproducible methodology encompassing measurements, analytical models and a utility-based method to design timelines filters. Using Facebook as a case study, our empirical results indicate that a significant bias exists and it is stronger at the topmost position of News Feed motivating the proposal of a novel and transparent fairness-based timeline design which can be controlled by users. Among the implications, we indicate the accuracy of our model to make counterfactual analysis, the capability of auditing social media and its versatility in designing multiple filters accounting for users preferences in an open and transparent way.

Maximizing Fair Content Spread via Edge Suggestion in Social Networks

“…Machine bias has recently been a focal concern in the social aspects of computer science and (big) data technologies [4,9], triggering extensive effort, including data pre-processing techniques [15,30,50,75], algorithm or model modification [52,96,97], and output post-processing techniques [46,51,90] to address these concerns. Social networks are not an exception when it comes to unfairness issues [47,48,68].…”