Hypothesis-margin Model Incorporating Structure Information for Feature Selection

Iterative search margin based algorithm (Simba) has been proven effective for feature selection. However, the previously proposed model does not effectively utilize the structure information hidden in data which may have a great impact on the generalization performance of post-analysis classifiers. In this paper, we introduce a novel hypothesis-margin model incorporating structure information for feature selection(Ssimba_FS). In the newly developed model, the structure information induced by clustering algorithms is incorporated into the existing hypothesis margin model for feature selection, and meanwhile the contribution of the structure information can be effectively adjusted by a trade-off parameter. Based on Ssimba_FS, we present a novel algorithm for feature selection (Ssimba). By Ssimba, an effectively ranked feature list can be obtained, futher a compact and relevant feature subset can be directly generated from the ranked feature list. The experiments on 6 real-life benchmark datasets show that the classifiers induced by the algorithm of this paper has better or comparable classification performance than those established by Simba in most cases.