Exact Linear-Time Algorithm for Parameterized K-Means Problem with Optimized Number of Clusters in the 1D Case

Abstract

We consider a well-known strongly NP-hard K-Means problem. In this problem, one needs to partition a finite set of N points in Euclidean space into K non-empty clusters minimizing the sum over all clusters of the intracluster sums of the squared distances between the elements of each cluster and its centers. The cluster’s center is defined as the centroid (geometrical center). We analyze the polynomial-solvable one-dimensional case of the problem and propose a novel parameterized approach to this case. Within the framework of this approach, we, firstly, introduce a new parameterized formulation of the problem for this case and, secondly, we show that our approach and proposed algorithm allows one to find an optimal input data partition and, contrary to existing approaches and algorithms, simultaneously find an optimal clusters number in $\mathcal {O}(N)$ time.