Enhanced dynamic self-organizing maps for data cluster

In the algorithm of Kohonen's Self-Organizing Maps (SOM) at the beginning of cluster, the number of input vectors in training set has to be settled down, which leads to the bad flexibility and is against the unsupervised principle. Also the fixed output network structure will lead to over-use or lack-of-use to the neuron node. To improve the exist defect of SOM and at the same time keep its advantages, an enhanced dynamic self organizing maps algorithm is proposed. This new method based on the idea of classical Growing Hierarchical Self-organizing Map (GHSOM), take advantage of GHSOM's feature of self determine the structure reflects the variability of data. By put forward a new cycle network structure EDSOM overcome the limitation of neutron under-utilize and over-utilize caused by the boundary effect. The experiment of intrusion detection proved the efficiency of the algorithm. © 2013 Asian Network for Scientific Information.