Feature selection has become a heated research issue for datasets that contain large numbers of features and has the ability to remove irrelevant and redundant features and improve classification accuracy in an effective fashion. A feature selection algorithm based on relationship between attributes named NCMIPV is proposed to acquire the optimized feature subset based on the analysis of existing relevant literatures, and the performance of NCMIPV on UCI datasets is evaluated. Experiment results show that compared with original datasets, this algorithm tends to shrink the dimension of feature space effectively in a comparatively shorter length of time. Moreover, the misclassification rate appears to rival other algorithms. Overall performance of the proposed algorithm is obviously superior to its counterparts in certain situation.