An Investigation of Representations and Operators for Evolutionary Data Clustering with a Variable Number of Clusters

This paper analyses the properties of four alternative representation/operator combinations suitable for data clustering algorithms that keep the number of clusters variable. These representations are investigated in the context of their performance when used in a multiobjective evolutionary clustering algorithm (MOCK), which we have described previously. To shed light on the resulting performance differences observed, we consider the relative size of the search space and heuristic bias inherent to each representation, as well as its locality and heritability under the associated variation operators. We find that the representation that performs worst when a random initialization is employed, is nevertheless the best overall performer given the heuristic initialization normally used in MOCK. This suggests there are strong interaction effects between initialization, representation and operators in this problem.