Learning similarity measure of nominal features in CBR classifiers

Abstract

Nominal feature is one type of symbolic features, whose feature values are completely unordered. The most often used existing similarity metrics for symbolic features is the Hamming metric, where similarity computation is coarse-grained and may affect the performance of case retrieval and then the classification accuracy. This paper presents a GA-based approach for learning similarity measure of nominal features for CBR classifiers. Based on the learned similarities, the classification accuracy can be improved, and the importance of each nominal feature can be analyzed to enhance the understanding of the used data sets. © Springer-Verlag Berlin Heidelberg 2005.