On Using Dimensionality Reduction Schemes to Optimize Dissimilarity-Based Classifiers

摘要

The aim of this paper is to present a strategy by which a new philosophy for pattern classification pertaining to dissimilarity-based classifiers (DBCs) can be efficiently implemented. Proposed by Duin and his co-authors, DBCs are a way of defining classifiers among classes; they are not based on the feature measurements of individual patterns, but rather on a suitable dissimilarity measure among the patterns. The problem with this strategy is that we need to select a representative set of data that is both compact and capable of representing the entire data set. However, it is difficult to find the optimal number of prototypes and, furthermore, selecting prototype stage may potentially lose some useful information for discrimination. To avoid these problems, in this paper, we propose an alternative approach where we use all available samples from the training set as prototypes and subsequently apply dimensionality reduction schemes. That is, we prefer not to directly select the representative prototypes from the training samples; rather, we use a dimensionality reduction scheme after computing the dissimilarity matrix with the entire training samples. Our experimental results demonstrate that the proposed mechanism can improve the classification accuracy of conventional approaches for two real-life benchmark databases.