Density Function Interpretation of Subspace Classification Methods

Subspace methods are a powerful class of statistical pattern classiication algorithms. They are traditionally considered nonparametric or semiparametric models, which means that the classiier does not produce the a posteriori probability of the input vector. This paper, however, presents a novel density estimation interpretation for the subspace methods. As a byproduct, the problem of unequal priors, which occurs if standard subspace methods are used when the a priori probabilities of the classes are unequal, is eliminated. As well, the a posteriori probability produced by the suggested method can be used to implement a plausible rejection method for diicultly classiiable patterns. At the end of this presentation, the applicability of the proposed probability density function interpretation is demonstrated with experiments.