On Undecided LP, Clustering and Active Learning

Abstract

We study colored coverage and clustering problems. Here, we are given a colored point set, where the points are covered by k (unknown) clusters, which are monochromatic (i.e., all the points covered by the same cluster have the same color). The access to the colors of the points (or even the points themselves) is provided indirectly via various oracle queries (such as nearest neighbor, or separation queries). We show that one can correctly deduce the color of all the points (i.e., compute a monochromatic clustering of the points) using a polylogarithmic number of queries, if the number of clusters is a constant. We investigate several variants of this problem, including Undecided Linear Programming and covering of points by k monochromatic balls. 2012 ACM Subject Classification Theory of computation → Computational geometry