Sparse Fisher discriminant analysis with Jeffrey's hyperprior

Abstract

The penalty function requires a choice of regularization parameter which controls the degree of parsimony in sparse kernel classifier. This involves an extra parameter apart from kernel parameter in the optimization which must be found via, e.g. cross-validation. This paper introduces a new parsimonious binary kernel Fisher discriminant analysis which does not require a regularization parameter. This can be done by using a Jeffrey's noninformative hyperprior. A Jeffrey's noninformative hyperprior is parameter-free and is adopted through a hierarchical-Bayes interpretation of the Laplacian prior distribution. This leads to a non-requirement of the regularization parameter. The proposed algorithm is compared with other machine learning methods on substantial benchmarks. Moreover, it is also compared with the leading machine learning in virtual screening application. It is found to be less accurate but it is still comparable in a number of cases.