On Some Variants for Anticlustering

Anticlustering involves partitioning objects into groups such that intergroup similarity is high and intragroup heterogeneity is high. In this paper, we propose five methods for anticlustering. The first proposed method minimizes the distances between group means. The second method minimizes both the distances between group means and those among group variances. The remaining three methods minimize the divergence among the distributions of groups using the Kullback-Leibler divergence, Jeffreys divergence, and Wasserstein distance. Through numerical experiments using an artificial dataset, the latter three methods prove superior to the others and the two conventional methods in terms of anticlustering quality and the perspective of parameter setting.