Authors - Libero Nigro, Franco Cicirelli Abstract - This paper builds on the Hartigan-Wong (HW) algorithm for unsupervised clustering. Although basic HW comes with an intrinsic high computational cost, it is known to be a better solution than K-Means, because it is less likely to get stuck around a sub-optimal solution of the data space. The paper, in particular, proposes a variant of HW, named Evolutionary HW (E-HW), which embodies genetic concepts and favors the achievement of more accurate clustering. E-HW depends on the use of a population of candidate solutions (centroid configurations), preliminarily created. E-HW is fed by a solution extracted from the population, which gets refined (crossed) and possibly replaced (mutated) following the basic HW operations. New generations of the population then come into existence. E-HW can be repeated a certain number of times, after that, experimental results highlight that the population favors the emergence of a solution close to the optimal one. To smooth out the computational burden, many operations of E-HW are implemented in parallel Java, so as to exploit the computing benefits of modern multi-core machines. The paper demonstrates the effectiveness of E-HW by using a collection of benchmark datasets, and the clustering results are compared with those achieved by competitor algorithms.
Open Zoom