Detection of locally relevant variables using SOM–NG algorithm

The SOM–NG algorithm is a combination of the Self-Organizing Map and the Neural Gas algorithms. It was developed to combine quantization and topological preservation. The algorithm also has a supervised version to create local linear models of scalar fields in the defined Voronoi regions. In this work a new methodology is proposed to use those models as data mining tools. Using visual tools, gradients are analysed to discover the influence of each variable over the output. It does not only allow to select the most relevant variables but also to detect different zones of influence, which can be used to create a set of fuzzy rules. The proposed methodology is proven to be useful to detect locally relevant variables that lead to a better understanding of the data.