Application of kohonen neural network and rough approximation for overlapping clusters optimization

Abstract

In this paper, the Kohonen Self Organizing Map one of the most popular tools in the exploratory phase of pattern recognition is proposed for clustering the input data. Recently researchers found that to have precise and optimized clustering operations and also to capture the ambiguity that comes from the data sets, it is not necessary to have crisp boundaries in some clustering operation. To overcome the mentioned ambiguity, two variation of cluster approximation (upper and lower) have been applied by Rough set theory. In the first stage the SOM is employed to produce the prototypes then, in the second stage the rough set is applied on the output grid of SOM to detect the ambiguity of SOM clustering. One of the most general optimization techniques (Simulated Annealing) have been adopted to assign the overlapped data to true clusters they belong to by minimizing the uncertainty criteria. Experiments show that the proposed two-level algorithm is more accurate and generates fewer errors as compared with crisp clustering operations.