A New Multidimensional Feature Transformation for Linear Classifiers and Its Applications

Abstract

In this paper, a new feature transformation method is introduced to decrease misclassification rate. Linear classifiers in general are not able to classify feature vectors which lie in a high dimensional feature space. When the feature vectors from difference classes have underlying distributions which are severely overlapped, it is even more difficult to classify those feature vectors with desirable performance. In this case, data reduction or feature transformation typically finds a feature subspace in which feature vectors can be well separated. However, it is still not possible to overcome misclassifications which results from the overlapping area. The proposed feature transformation increases the dimension of a feature vector by combining other feature vectors in the same class and then follows typical data reduction process. Significantly improved separability in terms of linear classifiers is achieved through such a sequential process and is identified in the experimental results.