Genetic Algorithms in Partitional Clustering: A Comparison

Three approaches to partitional clustering using genetic algorithms (GA) are compared with k-means and the EM algorithm for three real world datasets (Iris, Glass and Vowel). The GA techniques differ in their encoding of the clustering problem using either a class id for each object (GAIE), medoids to assign objects to the class associated with the nearest medoid (GAME), or parameters for multivariate distributions that describe each cluster (GAPE). For the simple Iris dataset, all algorithms except GAIE obtained results with comparable accuracy, but k-means and EM had more runs with inferior results compared to GAME and GAPE. For the more complex Glass dataset, the results for GAME and GAPE were superior compared to k-means, EM and GAIE regarding their accuracy and variance of the results for repeated runs. None of the algorithms was superior for the most complex dataset (Vowel). We conclude that GAs in clustering are a valuable alternative to k-means and EM, but that the choice of the problem representation is crucial. Key-words: partitional cluster analysis, genetic algorithms, optimisation in clustering, k-means, Expectation Maximisation (EM).