Vector Gravitation Clustering Networks.

In pattern recognition, patterns are described in terms of features. The features form feature vectors in the feature space. In the light of the phenomenon of gravitation in star clusters, we define patterns in the feature space to self-organize into clustering networks called "vector gravitation clustering networks" in this study. In the proposed clustering method, one called "vector gravitational force" is employed for the similarity measure in the feature space. Then by means of the "vector gravitational force", patterns self-organize clustering networks called "vector gravitation clustering networks" in the feature space. The proposed clustering method is applied to experiments. The experimental results show workability of the proposed clustering method. It is revealed that patterns tend to have more called "vector gravitational force" between ones of the same categories than between ones of the different categories in the feature space. Finally, further performance analysis employing the ANOVA ("analysis of variance") and the Newman-Keul procedure indicates potentiality of the proposed clustering method. As being inspired by the phenomenon of gravitation in star clusters and by using the "vector gravitational force" for similarity measure, "interpretability" is one obvious advantage of the proposed clustering method, and it may be viewed as one natural clustering method. [ABSTRACT FROM AUTHOR]