A novel validity index in fuzzy clustering algorithm

Fuzzy clustering algorithms, especially the fuzzy c-means (FCM) algorithm, need the number of clusters specified manually in advance, which, to some extent, limits the application of FCM algorithm. To solve this problem of fuzzy clustering algorithm not being able to predict the number of clusters, this paper proposes a new cluster validity index. This algorithm produces membership matrix and cluster centroid by implementing the FCM algorithm iteratively, figuring out the corresponding inter-class degree of separation, intra-class compactness and inter-class degree of overlap. A new cluster validity index voscar is defined by using the three indices. This new index solves the problem that FCM algorithm needs to preset a number of clusters, as well as avoiding the low accuracy of fuzzy degree introduced in the calculation of distance due to uneven distribution of data, particularly in the presence of intra-class overlap. The experiments conducted on artificial data sets and the actual data sets show that the proposed index can make correct evaluations of the fuzzy clustering results, and can automatically obtain the optimal number of clusters so as to improve the effect of clustering. Meanwhile it also shows that this index possesses excellent reliability for different fuzzy factors. © 2016 Inderscience Enterprises Ltd.