Fusion of Over-Segmentations for Improved Data Clustering.

This paper presents a method called fusion of over-segmentations (FOOS) that gives the optimal or near-optimal clustering solution in just three runs of k-means. The algorithm over-segments the dataset twice. The two over-segmentations fuse together and determine the initial cluster-centres for the third and last run of k-means. FOOS gives a plausible solution even if an approximate number of clusters is known. Theoretical, analytical and experimental results compare FOOS with single-run k-means and k-means with multiple restarts. This technique addresses the issue of data clustering in all the possible areas, including data mining, pattern recognition, machine learning and statistics. The results obtained while segmenting a variety of texture images and non-image datasets are encouraging. The 40% over-segmentation gives the best results. [ABSTRACT FROM AUTHOR]