Discussion Of Improving Fuzzy K-means Clustering

Cluster analysis is an important branch in Statistics and applied in many fields, including artificial intelligence and pattern recognition, marketing, psychometrics, chemometrics, etc. There are a lot of different methods, and partitioning method becomes the most commonly used one due to its efficiency. The representative algorithms include k-means, k-medoids and their various variants.The original k - means algorithm is actually a kind of hard computing. Since fuzzy set theory which produced the idea of uncertainty of belonging described by a membership function was proposed in 1965, soft computing was applied in clustering and fuzzy clustering has been widely studied. The representative algorithms include FANNY, FKM and their variants.The effect of clustering algorithms relies on the dissimilarity measure. In section 4.3 we introduce a new dissimilarity measure, AE metric. Experiments show that the new metric is more robust than the Euclidean norm.In this paper we present four new algorithms, the alternative fuzzy k-means (AFKM), the alternative weighted fuzzy k-means(AWFKM), the further improvement of WFKM and AWFKM(IWFKM and IAWFKM), which modify the original algorithms mainly by two ways (1) using the new dissimilarity measure, AE metric, to replace the Euclidean norm, (2) relaxing the constraints to the membership coefficients. We illustrate the advantages of new algorithms with several examples.