The impact of learning parameters on Bayesian self-organizing maps: An empirical study

The Bayesian self-organizing map (BSOM) algorithm is an extended self-organizing learning process, which uses the neurons' estimated posterior probabilities to replace the distance measure and neighborhood function. It is used in such areas as data clustering and density estimation. However, the impact of learning parameters has not been rigorously studied. Based on the analysis of two synthetic datasets, this paper investigates the impact of the selection of learning parameters such as the learning rates, the initial mean values, the initial covariance matrices, the input order and the number of iterations. The experimental results indicate that the BSOM algorithm is not sensitive to the initial mean values and the number of iterations, however, it is rather sensitive to the learning rates, the initial covariance matrices and the input order.