Themis: Modeling, Measuring and Mitigating Bias in Online Information Platforms

Themis: Modeling, Measuring and Mitigating Bias in Online Information Platforms

We live in a world where most of our information and communication needs are satisfied by online information platforms (OIPs) such as Search Engines and Social Networks and Media. These platforms play an important role in shaping the opinions and guiding the decisions of users for simple or important matters. Their operation relies on sophisticated machine learning and AI algorithms for filtering, ranking and recommendations, trained on massive amounts of data collected by the behavior and the contributions of users online. However, the function of these platforms can be compromised due to the presence of bias in the behavior of human users, but also in the decisions of the automated AI algorithms used by OIPs. These biases can result in unfair decisions by the AI algorithms, in recommendations of content, ranking of results, or the profiling of the users, as well as the emergence of echo chambers and filter bubbles in social networks.

The goal of the Themis project is to develop a formal framework for modeling, measuring, and mitigating bias online.

Themis will consider different types of bias in online information providers such as search engines and social networks and media. It will provide novel definitions of bias and fairness for problems such as opinion formation and community detection, and provide models for the emergence of bias in social media. It will also perform measurements of bias on datasets. The project is structured along three axes aligned with the three goals above.