Hybrid Dissimilarity Measurement for Intelligent Weight K-means Clustering

This paper represents one of the steps in overcoming the drawback of Weight K-means (WK-means), without considering the direction with respect to feature weights in feature space, using hybrid dissimilarity measure. This paper is aimed at breaking this point and facilitating feature space transformations directed by an intelligent Minkowski Weight K-means clustering algorithm through hybrid dissimilarity measure (iMWK-HD). The proposed hybrid dissimilarity measure assign weights to features based on Minkowski distance and Cosine dissimilarity, which can extend, shrink and rotate feature space so that the performance of WK-means clustering can be improved. In iMWK-HD, a new optimization objective function is designed based on minimizing the hybrid dissimilarity. Using the new objective function, it derives new updating rules for the iterations in the clustering procedure. Experimental results on UCI datasets demonstrate that iMWK-HD is superior to the three existing clustering algorithms, i.e. iK-means, iWK-means and iMWK-means. In addition, the proposed algorithms are immune to irrelevant features in cluster subspace.