Simulation Study for Cluster Analysis

Abstract

A Monte Carlo simulation study of four clustering methods -Complete linkage

, Ward's method, K-means method and AutoClass Bayesian classification-for determining the best partitioning is conducted on some artificial data sets. The results indicated that the classification obtained by the four clustering methods is affected by increasing the number of clusters in the data sets when error-free data sets are used. All clustering methods perform better than AutoClass Bayesian classification method when number of clusters in data sets is small while the performance of AutoClass Bayesian classification is better than other clustering methods when number of clusters in data sets is large. The results indicated that the classification obtained by the four methods is affected by dimension, noise, and outliers conditions. All clustering methods are affected by outliers. The performance of all clustering methods is decreased with increasing the number of clusters except for AutoClass Bayesian classification, its performance is increased with increasing the number of clusters.