Fair and Representative Clustering for Data Summarization

Abstract

Clustering is widely used in data summarization systems such as web search and personalization to produce representations consumed by society at scale. These systems typically partition data into k clusters following an objective function, and then select one representative from each cluster. Under current fairness notions, a data summary is deemed fair if it contains a sufficient number of representatives, or cluster centers, from each demographic group. I argue that such criteria are insufficient to ensure that a clustering is representative and fail to account for the context underlying data summarization applications. To address this, I propose a more expressive fairness notion for clustering, for which I develop approximation algorithms with provable representation guarantees.