A Novel Feature Selection Algorithm Based on Hypothesis-Margin

Iterative search margin based algorithm(Simba) has been proven effective for feature selection. However, it still has the following disadvantages: (1) the previously proposed model still lacks enough robust to noises; and (2) the given model does not use any global information, in this way some useful discrimination information may be lost and the convergence speed is also influenced in some cases. In this paper, by incorporating global information, a novel margin based feature selection framework is introduced. According to the newly designed model, an improved margin based feature selection algorithm(Isimba) is proposed. By effectively adjusting the contribution of the global information, Isimba can efficiently reduce the computational cost and at the same time obtain more effective feature subsets as compared to Simba. The experiments on 6 artificial and 8 real-life benchmark datasets show that Isimba is effective and efficient.