Feature selection by picking a small size of important features out of the feature space facilitates learning algorithms to perform more accurately and more efficiently on the datasets. Considering the universal existence of relevance between features in real datasets, this paper proposes an unsupervised feature selection framework in which the feature correlating to each other form a network structure and the importance of each of them is measured by degree centrality index of a complex network. The bigger the degree centrality of a feature in this network, the higher the rank of its importance. At the end we select a given number of features with the highest ranks. This framework allows more flexibility on handling feature importance and feature redundancy. Later the proposed method will be compared to classical selection/extraction techniques on six high?dimensional datasets. Experiments demonstrate the advantages of our model on both continuous and discrete datasets.