On resolution coresets for constrained clustering

TL;DR

This paper investigates resolution coresets for constrained clustering in materials science imaging, demonstrating that structured resolution coresets can outperform unrestricted coresets in bounds and computational efficiency.

Contribution

It introduces the concept of resolution coresets for weight-balanced clustering and proves bounds that improve upon existing results, enabling faster algorithms.

Findings

Resolution coresets can be more effective than unrestricted coresets.

Structured resolution coresets lead to improved bounds in clustering.

Algorithms using resolution coresets are significantly faster in practice.

Abstract

Specific data compression techniques, formalized by the concept of coresets, proved to be powerful for many optimization problems. In fact, while tightly controlling the approximation error, coresets may lead to significant speed up of the computations and hence allow to extend algorithms to much larger problem sizes. The present paper deals with a weight-balanced clustering problem from imaging in materials science. Here, the class of desired coresets is naturally confined to those which can be viewed as lowering the resolution of the data. Hence one would expect that such resolution coresets are inferior to unrestricted coreset. We show, however, that the restrictions are more than compensated by utilizing the underlying structure of the data. In particular, we prove bounds for resolution coresets which improve known bounds in the relevant dimensions and also lead to significantly faster algorithms practice.