Kernel method for percentile feature extraction

A method is proposed which computes a direction in a dataset such that a specified fraction of a particular class of all examples is separated from the overall mean by a maximal margin. The projector onto that direction can be used for class-specific feature extraction. The algorithm is carried out in a feature space associated with a support vector kernel function, hence it can be used to construct a large class of nonlinear feature extractors. In the particular case where there exists only one class, the method can be thought of as a robust form of principal component analysis, where instead of variance we maximize percentile thresholds. Finally, we generalize it to also include the possibility of specifying negative examples.