Multiple growing self-organizing map for data classification

Self-organizing map (SOM) is an unsupervised artificial neural network which is used for data visualization and dimensionality reduction purposes. Multiple self-organizing maps (MSOMs) enable the advantage of supervised learning scheme along with the unsupervised mapping. Same as SOMs, deciding the MSOM network structure at the initialization phase is a major limitation of this method. To work around this limitation, the growing self-organizing map (GSOM) which can visualize high-dimensional data using a dynamic network structure is employed. Here, we proposed a multiple growing self-organizing map (MGSOM) to benefit from both SOM and MSOM advantages alongside dynamic structure of GSOM. In the proposed method, feature vectors with the same classmembership are presented to independent GSOM networks. The proposed learning scheme improves the quality of learning and increases the classification performance significantly. The proposed method is compared to counter propagation artificial neural network (CPANN) and supervised Kohonen Network (SKN) on seven benchmark datasets and the results indicate the high ability of MGSOM to solve pattern classification problems.

Self-organizing map (SOM) is an unsupervised artificial neural network which is used for data visualization and dimensionality reduction purposes. Multiple self-organizing maps (MSOMs) enable the advantage of supervised learning scheme along with the unsupervised mapping. Same as SOMs, deciding the MSOM network structure at the initialization phase is a major limitation of this method. To work around this limitation, the growing self-organizing map (GSOM) which can visualize high-dimensional data using a dynamic network structure is employed. Here, we proposed a multiple growing self-organizing map (MGSOM) to benefit from both SOM and MSOM advantages alongside dynamic structure of GSOM. In the proposed method, feature vectors with the same classmembership are presented to independent GSOM networks. The proposed learning scheme improves the quality of learning and increases the classification performance significantly. The proposed method is compared to counter propagation artificial neural network (CPANN) and supervised Kohonen Network (SKN) on seven benchmark datasets and the results indicate the high ability of MGSOM to solve pattern classification problems.