Genetic-algorithms-based approach to self-organizing feature map and its application in cluster analysis

In the traditional form of the self-organizing feature map (SOFM) algorithm, the criterion for stopping training is either to terminate the training procedure when no noticeable changes in the feature map are observed or to stop training when the number of iterations reaches a prespecific number. In this paper we propose an efficient method for measuring the degree of topology preservation. Based on the method we apply genetic algorithms (GAs) in two stages to form a topologically ordered feature map. We then use a special method to interpret a SOFM formed by the proposed GA-based method to estimate the number and the locations of clusters from a multidimensional data set without labeling information. Two data sets are used to illustrate the performance of the proposed methods.