Using unclassified observations for improving classifiers

Summary

Methodologies for updating a classifier using unclassified observations are discussed. The focus is on classifiers based on linear or quadratic discriminant analysis. A semi-supervised clustering based on the Gustafson-Kessel algorithm for fuzzy clustering is carried out for all data, both classified and unclassified observations. The resulting fuzzy means and covariance matrices are used to update the classifier. It has formerly been shown that this methodology can reduce the misclassification rate. In this paper a modified approach is suggested for situations with errors in the data for the unclassified objects. To handle such situations, a noise cluster is introduced in the cluster analysis, and dubious points are allocated to this cluster. The proposed modifications are tested on simulated data. The results indicate that the misclassification rates are lower than or at the same level as with the original updating procedure. Copyright (C) 2004 John Wiley Sons, Ltd.