Adaptations of Relief for continuous domains of bioinformatics

Abstract

Relief occupies a niche among feature selection methods for data classification. Filters are faster, wrappers are much slower. Relief is feature-set-aware, same as wrappers. However, it is thought being able to deselect only irrelevant, but not redundant features, same as filters. Iterative Reliefs seek to increase the separation margin between classes in the anisotropic space defined by weighted features. Reliefs for continuous domains are much less developed than for categorical domains. The paper discusses a number of adaptations for continuous spaces with Euclidean or Manhattan metric. The ability of Relief to detect redundant features is demonstrated. A dramatic reduction of the feature-set is achieved in a health diagnostics problem.