Adaptive possibilistic clustering

Abstract

In this paper a new possibilistic clustering algorithm is proposed, where certain critical parameters are dynamically adjusted, allowing for increased flexibility in uncovering the clustering structure of the data. The new algorithm requires only a crude overestimation of the number of clusters (instead of the actual number of them, as many other well-known algorithms require), and has - in principle - the ability to reduce this number to that of the clusters, that are naturally formed by the data. In addition, since the proposed clustering algorithm is a possibilistic one, it is expected that it will provide dense in data points regions as clusters. Experimental results, on both synthetic and real data sets, verify the previous conclusions.