An Improved Clustering Algorithm Based on Fuzzy Set and Rough Set Theories

摘要

Cluster analysis is one of important techniques in data mining. Recently fuzzy set and rough set theories have been incorporated in the framework of k-means to develop the fuzzy k-means (FKM), rough k-means (RKM) and fuzzy rough k-means algorithms (FRKM). But the above rough clustering algorithms determining the lower and upper approximations based on the absolute differences of the distances between the samples and the cluster centroids, which is inappropriate while the distances of different samples to the clusters aren't on comparable scales. And the above rough clustering algorithms all omit the influence of the negative regions, which deteriorates their clustering performances. To tackle these problems, an improved fuzzy rough clustering algorithm is proposed, which utilizes approximation regions as well as negative regions. The criterion of the lower/upper approximations and negative regions is designed based on the relative difference of the distances between the samples and the clusters, and then the computation of cluster centroids is formulated as the weighted combination of the approximation regions and negative regions. To differentiate the contributions of different attributes, one new attribute weighting schemes is proposed. Finally experimental results show the advantages of our new clustering algorithm over the former clustering algorithm based on rough set.