Fair Clustering with Multiple Colors

TL;DR

This paper introduces a simple reduction method that achieves the first true constant factor approximation for fair clustering with multiple colors across various clustering objectives, improving fairness without significant loss in quality.

Contribution

It presents a novel reduction from unconstrained to fair k-clustering, enabling constant factor approximations for multiple clustering objectives with multiple colors.

Findings

First true constant factor approximation for multi-color fair clustering.

Reduction applies to k-median, k-means, and k-center objectives.

Achieves fairness with minimal impact on clustering quality.

Abstract

A fair clustering instance is given a data set $A$ in which every point is assigned some color. Colors correspond to various protected attributes such as sex, ethnicity, or age. A fair clustering is an instance where membership of points in a cluster is uncorrelated with the coloring of the points. Of particular interest is the case where all colors are equally represented. If we have exactly two colors, Chierrichetti, Kumar, Lattanzi and Vassilvitskii (NIPS 2017) showed that various $k$-clustering objectives admit a constant factor approximation. Since then, a number of follow up work has attempted to extend this result to a multi-color case, though so far, the only known results either result in no-constant factor approximation, apply only to special clustering objectives such as $k$-center, yield bicrititeria approximations, or require $k$ to be constant. In this paper, we present a simple reduction from unconstrained $k$-clustering to fair $k$-clustering for a large range of clustering objectives including $k$-median, $k$-means, and $k$-center. The reduction loses only a constant factor in the approximation guarantee, marking the first true constant factor approximation for many of these problems.