Check, Please: Verifiably Fair Clustering

TL;DR

This paper investigates the computational complexity of verifying proportional fairness in clustering, introduces new concepts for practical auditing, and provides algorithms for efficient verification.

Contribution

It proves coNP-hardness of verifying mPJR, introduces DC-mPJR+ for scalable auditing, and connects these concepts to practical clustering fairness guarantees.

Findings

Verifying mPJR is coNP-hard.

Introduces DC-mPJR+ with an $O(mn ext{ log } n + mnk)$ verification algorithm.

DC-mPJR+ serves as a practical proxy for global fairness in clustering.

Abstract

Popular centroid-based clustering methods are typically optimized for global objectives and may fail to adequately represent large groups of datapoints. To address this concern, recent work puts forward clustering analogs of social choice proportionality concepts, such as Proportionally Representative Fairness (also known as mPJR). For proportionality guarantees to be useful in practice, they must be (a) achievable and (b) efficiently auditable, so that one can check whether standard approaches, such as $k$-means, which are not guaranteed to provide proportional representation in general, nevertheless output proportional solutions on specific inputs. In this work, we study the computational complexity of verifying proportional representation in clustering. We first show that verifying mPJR is coNP-hard. Inspired by PJR+ -- a strengthening of PJR that is polynomial-time verifiable in the committee voting setting -- we introduce mPJR+ as its metric analog. However, verifying mPJR+ relies on repeated submodular minimization, rendering it impractical at scale. Hence, we introduce Default Coalitions mPJR+ (DC-mPJR+): a new proportionality concept that offers representation guarantees to a restricted set of coalitions around unselected centers, and as a result, admits an $O(mn \log n + mnk)$ verification algorithm. DC-mPJR+ is satisfied by SEAR and remains a meaningful proxy for global fairness: any solution satisfying $\gamma$-DC-mPJR+ also satisfies $(\gamma + 2)$-mPJR+. Together, our results identify a practical and theoretically grounded path for auditing proportional representation in clustering.