Abstract: A technique to visualize the trained self-organizing maps (SOM) networks results is discussed, and an approach for feature selection based on SOM networks combining with the visualization technique is presented. In the approach, the responsibilities of every dimensional feature in SOM networks competitive neurons weights to the input data are computed, and then the feature sets being, sensitive to pattern recognition, which have the main responsibilities, are selected accordingly. The experimental data sets including the well-known IRIS data and gearbox failure data are used to test the approach. It is proved from the investigation that efficient feature sets are chosen easily from raw feature sets, and then pattern recognition of input data is realized.

Key words: Feature selection, Pattern recognition, Self-organizing maps networks, Visualization