Approximate Clustering with Same-Cluster Queries

TL;DR

This paper introduces a polynomial-time approximation algorithm for the $k$-means clustering problem using a limited number of same-cluster queries, removing the need for margin assumptions and providing bounds on query complexity.

Contribution

It extends semi-supervised clustering with same-cluster queries to achieve $(1 + ext{epsilon})$-approximation without margin assumptions, using a query complexity independent of dataset size.

Findings

Achieves $(1 + ext{epsilon})$-approximation for $k$-means with few queries

Provides a lower bound on query complexity under ETH

Modifies $k$-means++ to obtain constant-factor approximation

Abstract

Ashtiani et al. proposed a Semi-Supervised Active Clustering framework (SSAC), where the learner is allowed to make adaptive queries to a domain expert. The queries are of the kind "do two given points belong to the same optimal cluster?" There are many clustering contexts where such same-cluster queries are feasible. Ashtiani et al. exhibited the power of such queries by showing that any instance of the $k$-means clustering problem, with additional margin assumption, can be solved efficiently if one is allowed $O(k^2 \log{k} + k \log{n})$ same-cluster queries. This is interesting since the $k$-means problem, even with the margin assumption, is $\mathsf{NP}$-hard. In this paper, we extend the work of Ashtiani et al. to the approximation setting showing that a few of such same-cluster queries enables one to get a polynomial-time $(1 + \varepsilon)$-approximation algorithm for the $k$-means problem without any margin assumption on the input dataset. Again, this is interesting since the $k$-means problem is $\mathsf{NP}$-hard to approximate within a factor $(1 + c)$ for a fixed constant $0 < c < 1$. The number of same-cluster queries used is $\textrm{poly}(k/\varepsilon)$ which is independent of the size $n$ of the dataset. Our algorithm is based on the $D^2$-sampling technique. We also give a conditional lower bound on the number of same-cluster queries showing that if the Exponential Time Hypothesis (ETH) holds, then any such efficient query algorithm needs to make $\Omega \left(\frac{k}{poly \log k} \right)$ same-cluster queries. Our algorithm can be extended for the case when the oracle is faulty. Another result we show with respect to the $k$-means++ seeding algorithm is that a small modification to the $k$-means++ seeding algorithm within the SSAC framework converts it to a constant factor approximation algorithm instead of the well known $O(\log{k})$-approximation algorithm.