The performance of CN-based similarity index is not satisfied due to only taking into account the local information of a network. The community information contains the network structure features of nodes, which can be adopted to improve the prediction accuracy. Therefore, a community-based link prediction algorithm using the community structure information is proposed to address the problem. Employing community relations and community information of common neighbors, it was developed in an attempt to improve the prediction precision. Firstly, two graph embedding methods--DeepWalk and Node2vec were employed, that is, a deep learning model, i. e. Skip-Gram was adopted to train the nodes’ sequences generated from short random walk and then the acquired embedding vectors of nodes were used in communities division to obtain high quality communities that contain more network topology information. Then, the similarity model of communities was proposed via defining the edge relationship between communities. Finally, the similarity of nodes, the similarity between the communities where the nodes are located, and the community information of the nodes’ common neighbors were integrated into the suggested algorithm to evaluate the link probability of two unknown nodes. Finally, experiments on six real-world networks like USAir are conducted, and the AUC of the suggested method is increased about 2. 3% at most compared with four benchmark algorithms including CN. Thus it shows that community structure information plays an important role when predicting the latent links.