Abstract:

To deal with the parameter sensitivity problem of possibilistic c-means clustering algorithm, a new possibilistic clustering algorithm based on center merging was proposed. The cluster number and structure were dynamically adjusted according to the data distribution. The algorithm has the ability to execute multi\|scale analysis task for the given data set by means of adjusting the values of the scale factor. The theorems were also given that were proven to be used to analyze the multiscale property of the algorithm. Compared with the traditional fuzzy or possibilistic clustering algorithms, the proposed algorithm avoids its dependence on the initial conditions of centers, cluster number and membership matrix, which makes it easy to control. Synthetic and real data experimental results show that the algorithm can be used to detect the cluster structures of the data set from different scales, and to find the clusters with different sizes.

Key words: possibilistic clustering, multiscale, center merging, initialization sensitivity