A method of data clustering based on improved algorithm of ART2

After such characteristics as normalization of vector and global vigilance parameter have been analyzed in the clustering process of classical Adaptive Resonance Theory Network (ART2), shortcomings of ART2 are pointed out, which are inapplicability to the situation correlative with vector modulus, inability of dividing space with different granularities according to the densities of space and output without hierarchical structure. So an improved algorithm of ART2 has been presented in this paper. This algorithm presents local vigilance parameter and pre-selection of neurons with standard of modulus of vector, and obtains dynamic clustering structure with hierarchy structure that is correlative with modulus of vector by cycling. This algorithm also reduces the requirement of setting vigilance parameter, that is, retraining neural network is not needed for clustering with bigger granularity. Such advantages can effectively satisfy basic demands of clustering and be adapted to the environment of clustering of sub-spaces respectively with characteristic of modulus.