INTERVAL TYPE-2 APPROACH TO KERNEL POSSIBILISCTIC C-MEANS CLUSTERING

Abstract

Pattern recognition and machine leaning is used to solve many complex problems of data mining, decision making and grouping of natural data. Among pattern recognition algorithms clustering is very popular and widely being used for problems of classification of image data as well as other real data. Fuzzy clustering replaced hard clustering due to its more robust clustering results and flexible classification boundary. Kernel fuzzy clustering is extensively used for clustering of nonlinear separable, complex distributed and ring shape like data sets, and moreover it is useful for data sets containing some sort of noisy pattern with random distribution. Kernel function is commonly defined as dot product of two values obtained by linear or nonlinear transfer function and has positive semi definite nature. Kernel trick adds additional degree of freedom by implicitly mapping of input patterns into higher dimensional space known as kernel space. The outcome of Kernel clustering is much better as compared to conventional fuzzy clustering algorithms such as fuzzy c-means, possibilistic c-means and possibilistic fuzzy c-mean not only for spherical data sets but also non spherical data sets. However in case of kernel possibilistic C-means, cluster coincidence occurs resulting into poor trained network. In this dissertation, we propose an Interval type 2 kernel possibilistic c-means to overcome this cluster coincidence problem in kernel possibilistic c-means clustering algorithm since type-2 fuzzy sets handle uncertainties better than type-1 fuzzy sets. Choice of kernel function is mostly data dependent, we selected Gaussian kernel with variable sigma for experiments. For same value of sigma of Gaussian kernel our proposed method outperforms and presented results will show the validity of our proposed method