Feature selection in machine learning via variable neighborhood search

The generalization ability of machine learning methods can be improved via feature selection. In this work a novel heuristic framework for feature selection in machine learning is proposed. The framework is built on the Variable Neighborhood Search (VNS) heuristic. The proposed framework is generic, and can be applied to any existing supervised machine learning methods. Implementation of the proposed framework that encapsulates conventional regression and classification problems is illustrated in this paper. Numerical experiments with real datasets display the applicability of the proposed framework.