A Novel Approach for Evaluating Class Structure Ambiguity

It is attractive and worthy to estimate the ambiguity of one existing class structure such that one could give suggestions to domain experts when and how to reorganize the original class structure. In this paper Class Structure Ambiguity (CSA) was proposed to estimate the quality of one class structure. To inspect whether the CSA did tell the quality of class structure or not, the Pearsons correlation between classification accuracies achieved by a linear SVM classifier and the values of CSA were evaluated according to two types of datasets, one generated randomly and another selected from the LIBSVM. The experimental results showed that the CSA did reveal the degree of the ambiguities among classes. To our knowledge, we were the first to discuss the problem of class structure ambiguity.