Comparison of Clustering Techniques for Cluster Analysis

Abstract

Cluster analysis is important for analyzing the number of clusters of natural data in several domains. Various clustering methods have been proposed. However, it is very difficult to choose the method best suited to the type of data. Therefore, the objective of this research was to compare the effectiveness of five clustering techniques with multivariate data. The techniques were: hierarchical clustering method; K-means clustering algorithm; Kohonen’s Self-Organizing Maps method (SOM); K-medoids method; and K-medoids method integrated with Dynamic Time Warping distance measure (DTW). To evaluate these five techniques, the root mean square standard deviation (RMSSTD) and r² (RS) were used. For RMSSTD, a lower value indicates a better technique and for RS, a higher value indicates a better technique. These approaches were evaluated using both real and simulated data which were multivariate normally distributed. Each dataset was generated by a Monte Carlo technique with 100 sample sizes and repeated 1,000 times for 3, 5 and 7 variables. In this research, 2, 3, 4, 5, 6, 7 and 8 clusters were studied. Both real and simulated datasets provided the same result, with the K-means clustering method having the closest RMSSTD and RS results to the SOM method. These two methods
yielded the lowest RMSSTD and highest RS in all simulations. Hence, both K-means and SOM were considered to be the most suitable techniques for cluster analysis.