Links prediction based on supervised learning is a main research topic in the field of complex network analysis. The core process of these methods is that the network is divided into training and target sets, then a classification model is used to learn the training set and forecast the missing links in target set. Such methods have two major challenges: first, we need to dig deep network information to define a set of features; Second, how to incorporate feature selection model to mine discriminative features. To solve the above problem, a model which integrates community features and mRMR feature selection was proposed. Such model first discovered global features associated with the link through the community, then used classical mRMR algorith-m metrics to measure the correlation between features, and filter out the best representative candidates by clearing noisy information. Experimental results show our proposed model can effectively improve the performance of link classification.