Cost-Sensitive Feature Selection using Particle Swarm Optimization: A Post-Processing Approach

Abstract

Feature selection is one of the important tasks in machine learning. Feature selection task deals with selecting a subset of feature from an original feature set. An important consideration in feature selection is the usefulness of a feature i.e. a set of feature which is selected is neither irrelevant nor redundant. Most of the existing algorithms in the domain of feature selection are designed to optimize the aforementioned objective. In our research we have addressed a third dimension of usefulness i.e. cost of the feature. Cost-effectives of a solution is most apt in cases where there is an asymmetric cost of data acquisition such as medical diagnosis applications. In this regard, our research deals with enhancing the existing feature selection techniques with a post-processing stage in which cost consideration is also accounted for. The resultant solution is optimized for both important as well as cost-effective features. We have used particle swarm optimization with post processing over chronic kidney disease dataset for generating a feature subset set which is both salient and cost-effective.