On Using Class-dependent Principle Component Analysis for Dissimilarity-Based Classifications

The aim of this paper 1 is to present an empirical evaluation on using class-dependent principle component analysis (PCA) for dissimilarity-based classifications (DBC) [18]. In DBC, the classification performance relies heavily on how well the dissimilarity space is constructed. In this paper, we study a way of constructing it in eigenspaces, spanned by the subset of principal eigenvectors, extracted from the training data set through the class-dependent PCA, instead of utilizing prototype selection methods and/or generalizing dissimilarity measures. In particular, we generate an eigenspace (i.e., a covariance matrix) per class, not for the entire data set, to compute distances in a vector space constructing a dissimilarity-based classifier. Our experimental results, obtained with well-known benchmark data and some UCI data sets, demonstrate that when the dimensionality of the eigenspaces has been appropriately selected, DBC, albeit not always, can be improved in terms of classification accuracies.