ISODATA: Thresholds for splitting clusters

Abstract

The author has identified the following significant results. The parameter AD (average distance) as used in the ISODATA program was critically examined. Thresholds of AD to decide on the splitting of clusters were obtained. For the univariate case, 0.84 was established as a sound choice, after examining several simple, as well as composite, distributions and also after investigating the probability of misclassification when points have to be reassigned to the newly identified clusters. For the multivariate case, the empirical threshold (N-0.16)/square root of N was extrapolated. A final criticism on AD was that AD would lose its effectiveness as a discriminative measure for the present purpose when N was large.