BAYESIAN SELF-ORGANIZING MAP FOR DATA CLASSIFICATION AND CLUSTERING

Abstract

The Bayesian self-organizing map (BSOM) has typically been used for density estimation. In this study, we implemented an adaptation of the model for performing unsupervized and supervised classification. In order to determine the optimal number of neurons to represent the given dataset during the learning process, an extended Bayesian learning process is proposed called the growing BSOM. It starts with two neurons and adds new neurons to the network via a process in which the neuron with the lowest individual log-likelihood is identified. The system has been tested using three synthetic datasets and one real dataset. The experimental results suggest that the BSOM-based approach can achieve better classification performance in comparisons to several widely-used models such as k-nearest neighbor (KNN), support vector machine (SVM) and Gaussian mixture model (GMM). By using the Bayesian information criterion (BIC) as a stopping criterion, the growing BSOM can model the data under study and estimate the number of clusters.