A new greedy randomized procedure for the feature selection problem

Feature Selection (FS) arises in data analysis to reduce the dimension of large data. We focus on integer programs to model FS problems, describe a variant with its mathematical properties, and design an ad-hoc solution strategy, whose performances are compared with those of well known FS methods. Our method is suitable for high-dimensional data instead of an exact solution approach. The experiments use randomly generated datasets and are designed to show the efficacy of both the FS model and of the heuristics. Finally, our method has been successfully applied to real biological problems.