Generalized Tree-Like Self-Organizing Neural Networks with Dynamically Defined Neighborhood for Cluster Analysis

摘要

The paper presents a generalization of self-organizing neural networks of spanning-tree-like structures and with dynamically defined neighborhood (SONNs with DDN, for short) for complex cluster-analysis problems. Our approach works in a fully-unsupervised way, i.e., it operates on unlabelled data and it does not require to predefine the number of clusters in a given data set. The generalized SONNs with DDN, in the course of learning, are able to disconnect their neuron structures into sub-structures and to reconnect some of them again as well as to adjust the overall number of neurons in the system. These features enable them to detect data clusters of virtually any shape and density including both volumetric ones and thin, shell-like ones. Moreover, the neurons in particular sub-networks create multi-point prototypes of the corresponding clusters. The operation of our approach has been tested using several diversified synthetic data sets and two benchmark data sets yielding very good results.