Ensemble Global k-Means via Clustering of Local Centroids: An Empirical Assessment

Undergraduate Research Project Report

This report documents an undergraduate research project conducted by Fatemeh Gholami Ghahdarijani under the supervision of Dr. Najmeh Hosseini Monjezi between September 2024 and July 2025.

The project investigated an ensemble approach to Global K-means initialization. The method partitions a dataset into subsets, applies local K-means clustering, pools the resulting representatives, and uses Global K-means to obtain initial centers for clustering the complete dataset.

The report presents historical experimental comparisons with conventional K-means and variants of Global K-means on four UCI datasets.

The revised October 2026 version clarifies the behavior of the original implementation when the requested number of local clusters exceeded the number of observations within a partition. According to the researcher's recollection, the implementation returned the available observations as representatives in such cases. This behavior provides a possible explanation for the convergence of some reported results toward those of Global K-means as the number of clusters increased.

The original implementation is no longer available, and the historical experiments have not been independently reproduced. The experimental results are preserved as records of the undergraduate project, with these limitations explicitly acknowledged.

This document is an undergraduate research report and has not undergone peer review.