Feature reduction is an important concept which is used for reducing dimensions to decrease the computation complexity and time of classification. Since now many approaches have been proposed for solving this problem, but almost all of them just presented a fix output for each input dataset that some of them aren’t satisfied cases for classification. In this we proposed an approach as processing input dataset to increase accuracy rate of each feature extraction methods. First of all, a new concept called dispelling classes gradually (DCG) is proposed to increase separability of classes based on their labels. Next, this method is used to process input dataset of the feature reduction approaches to decrease the misclassification error rate of their outputs more than when output is achieved without any processing. In addition our method has a good quality to collate with noise based on adapting dataset with feature reduction approaches. In the result part, two conditions (With process and without that) are compared to support our idea by using some of UCI datasets.