Data Clustering with Cluster Size Constraints Using a Modified K-Means Algorithm

Data clustering is a frequently used technique in finance, computer science, and engineering. In most of the applications, cluster sizes are either constrained to particular values or available as prior knowledge. Unfortunately, traditional clustering methods cannot impose constrains on cluster sizes. In this paper, we propose some vital modifications to the standard k-means algorithm such that it can incorporate size constraints for each cluster separately. The modified k-means algorithm can be used to obtain clusters in preferred sizes. A potential application would be obtaining clusters with equal cluster size. Moreover, the modified algorithm makes use of prior knowledge of the given data set for selectively initializing the cluster centroids which helps escaping from local minima. Simulation results on multidimensional data demonstrate that the k-means algorithm with the proposed modifications can fulfill cluster size constraints and lead to more accurate and robust results.