Semi_Fisher Score: A semi-supervised method for feature selection

Feature selection is an important problem for pattern classifier systems. As compared to unsupervised feature selection methods, supervised feature selection approaches have better performance when the given training samples with supervised information are sufficient. However, in reality, usually only a few labeled data are obtained, since obtaining class labels is expensive but many unlabeled data can be easily gotten. For this case, directly using the existing supervised feature selection algorithms may be failed because the data distribution may not be accurately estimated only by using a few labeled data. So, in this paper, we introduce a semi-supervised method for feature selection, called Semi_Fisher Score, the new model attempts to effectively simultaneously utilize all labeled and unlabeled samples for improving the performance of the classical Fisher Score. Experiments on 4 UCI datasets by using three different classifiers(KNN, RBFNN and C4.5)show the effectiveness of our algorithm.