Principal Directions-Based Algorithm for Classification Tasks

In our approach, we consider a probabilistic class model where each class h isin H is represented by a probability density function defined on Rⁿ where n is the dimension of input data and H stands for a given finite set of classes. The classes are learned by the algorithm using the information contained by samples randomly generated from them. The learning process is based on the set of class skeletons, where the class skeleton is represented by the principal axes estimated from data. Basically, for each new sample, the recognition algorithm classifies it in the class whose skeleton is the "nearest" to this example. For each new sample allotted to a class, the class characteristics are re-computed using a first order approximation technique. We introduce two principal directions based learning algorithms, a non-adaptive variant and an adaptive variant respectively. Comparative analysis is performed and experimentally derived conclusions concerning the performance of the new proposed methods are reported in the final section of the paper.