Monte Carlo comparisons of selected clustering procedures

A comparison of the classification capabilities of the 1-dimensional kohonen neural network with two pratitioning and three hierarchical cluster analysis algorithms

“…In contrast to our previous findings, the k-means (W) and k-means (A) missclassification rates are highly similar despite the fact that the Ward's and group average rates are noticeably different. Apparently, Hartigan's k-means algorithm can recover from poor starting values when the data are highly structured (see Bayne, Beauchamp, Begovich & Kane, 1980;Milligan, 1980). We noted earlier that Milligan's cluster generation method often produces highly structured data with indicator validities that are greater than .90.…”

The Methods and Problems of Cluster Analysis

“…Very few Monte Carlo studies have been performed which have used clustering methods from any of the other families of clustering techniques described earlier (cf. Bayne et a!., 1980;Boake, 1983). Also there has been relatively little comparison of the various similarity measures and their effects on the empirical performance of clustering techniques (cf.…”

“…A rational selection of the starting partition did little to improve this situation, but Milligan (1980) has demonstrated that the k-means pass, using an initial starting partition derived from average linkage clustering, provided superior recovery of known data structure when compared with the performance of other iterative and hierarchical clustering methods. Others have suggested that iterative methods produce optimal solutions regardless of the starting partition if the data are well structured (Everitt, 1980;Bayne, Beauchamp, Begovich, and Kane, 1980).…”

Comparing methods of classifying life courses: sequence analysis and latent class analysis

“…However, in the context of SA, we do not directly observe these underlying variables but only a diffuse summary measure: a distance between sequences. Bayne, Beauchamp, Begovich and Kane (1980), using bivariate distributions, tested thirteen different techniques for their classification accuracy. Their last, concluding sentence is "However, as the complexity of the distributions increases, the differences between all of these methods decrease".…”