In pattern recognition, knowledge of the structure of pattern data can help us to know the sufficiency of features and to design classifiers. We have proposed a graphical visualization method where the structure and separability of classes in the original feature space are almost correctly preserved. This method is especially effective for knowing which groups of classes are close and which groups are not. In this paper, we propose a method to group classes on the basis of this graphical analysis. Such a grouping is exploited to design a decision tree in which a sample is classified into groups of classes at each node with a different feature subset, and is further divided into smaller groups, and finally reaches at one of the leaves consisting of single classes. The main characteristics of this method is to use different feature subsets in different nodes. This way is most effective to solve multi-class problems. An experiment with 30 characters (30 classes) was conducted to demonstrate the effectiveness of the proposed method.
展开▼
