Supervised Kohonen network with heterogeneous value difference metric for both numeric and categorical inputs

The multi-attribute information appears in real world, which also includes numeric and categorical attributes. However, the previous classification algorithms for both numeric and categorical data exist in some limitations on categorical data. In this paper, a supervised Kohonen network with heterogeneous value difference metric is proposed for both numeric and categorical inputs. It employs the framework of supervised Kohonen networks, adopts heterogeneous value difference metric to measure dissimilarity between numeric and categorical data, uses the frequency of each categorical item in the Voronoi set to update the reference vector of categorical attribute on the competitive layer, and updates different competitive learning rules for numeric and categorical data. The effectiveness of the proposed algorithm is verified by UCI Machine Learning Data Repository. The classification accuracy is compared with BP, k-NN, naive Bayes network, C4.5 and SVM; the dissimilarity metric is analyzed. The proposed classification algorithm is applied to the operating mode classification for wind turbines; the effectiveness is illustrated in condition monitoring for pitch system of wind turbines.