A Better Clustering Validity Measure

摘要

Although the Xie-Beni clustering validity measure is currently the best one, we modify it to better handle the cases where the minimum distance between cluster centers is not representative of their typical distances. The result is that our measure is equivalent to the Xie-Beni measure in cases where that measure works well and is better in cases where the minimum distance between clusters is too small relative to the other distances. Our new validity measure is tested versus the Xie-Beni by use of a new clustering algorithm that improves the well-known k-means algorithm. The test objective is to determine the optimal number K of clusters for a set of feature vectors and to expose a pathology of the Xie-Bene validity.