A modified sequential deep floating search algorithm for feature selection

With the exponential increase of the data scale, the problem of feature selection has been the focus in statistical pattern recognition. In this paper, a new modified forward deep floating searching algorithm (SDFFS) is proposed to select a feature subset of d features from the original candidate-set of D features (d <; D), which is an improvement of the state of the art SFFS algorithm. The SDFFS algorithm adds an additional step called `deep searching´ to check whether there exists a better k-subset than the one selected by the SFS step, which can not be searched in the SFFS algorithm. The algorithm is tested on eight real datasets including four microarray gene expression datasets and four UCI datasets. Four feature selection algorithms SFS, PlMr, SFFS and IFFS are used to compare with our proposed SDFFS algorithm. The experimental results show that the SDFFS performs better than others in acceptable computational cost in most of datasets, which indicates that our designed deep searching step works well.