Abstract: Aiming at the problem that the existing methods made use of the relative fragmentation of network information, and it was difficult to describe the relationship between link number and similarity score, we proposed a link prediction method in dynamic networks, which used time series combined from node similarities scores and link mumber to predict. Firstly, it predicted similarity scores of all time snapshots by community evolution. Secondly, nodes similarities were combined with real link number by binary time series model and the probability of link between all node pairs in next period was predicted. Finally, the test was carried out on a dataset forwarded by WeiboNetTweet microblog. The experimental results show that the method improves the prediction accuracy by at least 5%, proves the intrinsic relationship between community evolution and link prediction, and verifies the effectiveness of binary time series model.

Key words: link prediction, community evolution, time series, node similarity