An unorthodox and effective non-iterative procedure for spherical clusters is demonstrated in this paper. It uses natural Bézier functions to determine initial cluster locations using the content of the data. The natural Bernstein-Bézier functions are very robust in representing data through continuous functions in the application of functional data analysis. This paper demonstrates that they are equally robust at resolving data clusters in classification problems. The original data is scaled and segmented. A natural Bézier function is fitted for each segment and the initial clusters are centered at the function extremums that are distinctly located. A self-selection process based on least distance is used to assign the data to these initial cluster centers. A minimum membership count is imposed and nearby clusters are combined to reduce these initial cluster centers based on visual clues. Centroid recalculation and data reassignment can be used for centroid convergence. The method is demonstrated for two dimensional spherical clusters. This approach requires no iteration. Four examples from clustering benchmark datasets are used to showcase the method. The data include different numbers of clusters, different data density for the clusters, as well as different levels of overlap. This method is new and different from other data clustering methods available in the literature. It is better than the standard k-means clustering method since it does not require information on the number of clusters or cluster membership count. The method is non-iterative and does not require random initialization or distance optimization.

This content is only available via PDF.
You do not currently have access to this content.