A Projected Feature Selection Algorithm for Data Classification

In contrast to many popular feature selection algorithms that provide suboptimal solutions according to some criterion, the OCFS algorithm can ensure optimal solutions according to the orthogonal centroid criterion. Based on the properties of OCFS, this paper proposes a projected feature selection algorithm called projected OCFS (POCFS) for data classification. POCFS extends OCFS to select different features for each class pair individually rather than to select the same features for all the classes simultaneously. Thus, It can select more suitable features for classifier construction than OCFS. Experimental results on real data set KDD- CUP99 indicate that POCFS outperforms OCFS in terms of their effectiveness and efficiency.