Semi-random subspace sampling for classification

In this paper, we introduce a novel semi-random subspace sampling for classification (for short, denoted by FS_RS). In this method, a ranking feature list is obtained by using feature selection first, and then the more important N0 features in the front of the ranking feature list are chosen, and N1 features is randomly selected from the remaining features in the ranking feature list. Along this sampling method, those obtained feature subsets not only contain those more important features, but also include those relatively weak relevant or irrelevant features, hence both diversity and accuracy of corresponding base classifiers can be effectively guaranteed. So, the performance of the integrated classifier can be effectively improved. Experiments on 4 real-life datasets show the effectiveness of our method.