One of the challenges in inferring a classification model with good prediction accuracy is to select the relevant features that contribute to maximum predictive power. Many feature selection techniques have been proposed and studied in the past, but none so far claimed to be the best. In this paper, a novel and efficient feature selection method called Clustering Coefficients of Variation (CCV) is proposed. CCV is based on a very simple principle of variance-basis which finds an optimal balance between generalization and overfitting. Through a computer simulation experiment, 44 datasets with substantially large number of features are tested by CCV in comparison to four popular feature selection techniques. Results show that CCV outperformed them in all aspects of averaged performances and speed. By the simplicity of design it is anticipated that CCV will be a useful alternative of pre-processing method for classification especially with those datasets that are characterized by many features.