An Improved Method for K-Means Clustering

K-Means has been paid attention to many areas recently, however, it is easy to fall into local optimum and the outliers influence the final results. This paper proposes an improved method for k-means clustering. Different from the traditional k-means algorithms, in our algorithm both intracluster compactness and intercluster separation are considered in our new presented method. A new model is established for hard cluster assignments of k-means clustering. Our new method transforms the problem of clustering to an integer programming problem and genetic algorithm is introduced to update cluster assignments iteratively. Experiment results on UCI data sets have showed the potential performance improvement of our method.