The algorithms used to sort and filter social media feeds are generally opaque. The goals of those algorithms are not visible to the user and may not align with their preferences but rather the goals of the company who constructed them. In my thesis, I am developing a tool that allows users to have finer grained control over their social media feeds as well as the ability to audit the performance of the algorithms being used to produce those feeds. This introduces several computational questions that need to be addressed including: the method of combining output from multiple algorithms, efficiency for use at scale, usability of the system for non-expert users, and evaluation of the performance of the system. So far we have developed and released our aggregation tool, called Gobo which aggregates Mastodon, Bluesky, Reddit, and, if their API still allowed it, X, formerly known as Twitter. Before the end of 2023 we will release proof-of-concept releases that integrate simple algorithmic filtering and ranking into the tool, which already supports simple keyword and username exclusion. Our approach to recommendation is to combine rules based recommendation with scores from neural networks in order to surface posts that the user is interested in. Our approach to auditing is to create a tool that assists the user in selecting a diverse set of posts on which they can run the algorithm to see the results. The intent is to have an interface that allows users to specify sets of posts, provides reference sets, and suggests posts. In addition, said interface will allow the user to adjust some of the model parameters to see how that changes the scoring. In addition, we plan to outline an ecosystem for allowing third parties to construct, customize, and deploy these algorithms. The expected contributions of this work are a new social media aggregator, a transparent cross-platform social media recommendation system, a tool for interactively auditing the performance of the recommendation system, and an outline for the implementation of an ecosystem of tools for constructing, customizing, and deploying variations of the recommendation system. We expect that the recommendation system will be influential in the web and social media community. We expect that the auditability will be influential in the auditable and explainable AI community.
Advisor: Ethan Zuckerman