New hierarchical model using SOM improves accuracy of classifiers for multiclass data sets

It is our observation that a classification model for a multiclass problem creates a large number of rules in the model and reduces its accuracy, especially if the number of features is large. Many strategies have been outlined in the literature to reduce the multiclass problem into smaller tasks (mostly creating two classes) and combining the results of the same to obtain the complete classification. In this paper, we present a solution for reducing the misclassification rate of the rule-based model using hierarchical clustering. The method uses a two-step approach. First, Self Organizing Maps (SOM), a visual clustering technique is used to identify the number of clusters that are naturally formed. This number of clusters is used as an indication of the number of classes that reduce the misclassification rate. The second step is to use the concept of hierarchical classification, and One versus All (OVA) to systematically reduce the number of classes. These results are encouraging, with a significantly reduced misclassification rate along with reduced number of evaluations (number of classifier runs) and simplified decision trees.