A method of data clustering based on improved algorithm of ART2

Abstract

After such characteristics as competitive rules and vigilance parameter have been analyzed in the clustering process of classical adaptive resonance theory network (ART2), shortcomings of ART2 are pointed out, which are subjective setting of vigilance parameter, excessive dependence on winning neuron information and output without hierarchical structure. So an improved algorithm of ART2 has been presented in this paper. This algorithm makes use of single ART2 neural network to realize multi-layer dynamic clustering structure by simultaneously considering the information from both the winning neuron and other neurons in process of competition as well as Hebb rule. Therefore, retraining neural network is not needed for clustering with granularity of a certain range. In addition, this algorithm reduces the requirement of subjective setting of vigilance parameter. Such advantages can effectively satisfy basic demands of clustering and avoid the problems of performance and parameter setting in realizing hierarchical structure by adopting cascade structure.