27
2
v1v2 (latest)

Distributed k-Means with Outliers in General Metrics

Abstract

Center-based clustering is a pivotal primitive for unsupervised learning and data analysis. A popular variant is undoubtedly the k-means problem, which, given a set PP of points from a metric space and a parameter k<Pk<|P|, requires to determine a subset SS of kk centers minimizing the sum of all squared distances of points in PP from their closest center. A more general formulation, known as k-means with zz outliers, introduced to deal with noisy datasets, features a further parameter zz and allows up to zz points of PP (outliers) to be disregarded when computing the aforementioned sum. We present a distributed coreset-based 3-round approximation algorithm for k-means with zz outliers for general metric spaces, using MapReduce as a computational model. Our distributed algorithm requires sublinear local memory per reducer, and yields a solution whose approximation ratio is an additive term O(γ)O(\gamma) away from the one achievable by the best known sequential (possibly bicriteria) algorithm, where γ\gamma can be made arbitrarily small. An important feature of our algorithm is that it obliviously adapts to the intrinsic complexity of the dataset, captured by the doubling dimension DD of the metric space. To the best of our knowledge, no previous distributed approaches were able to attain similar quality-performance tradeoffs for general metrics.

View on arXiv
Comments on this paper