The problem of link prediction in complex network has attracted increasing attention recently. It could be used to find out missing links (recovery problem) or recommend new links (recommendation problem). Traditional algorithms (e.g., CN, RA) extract the missing information and predict new links based on the structure of network or the feature of node. These algorithms have an implicit hypothesis that the network tends to be completed. This is not always in accord with the reality. Most of them did experiments on un-weighted graphs. That means they treated all the links equally and ignored the effect of temporal information. What’s more some algorithms only focused on recovery problem. Their performance maybe not good at recommendation new links because the features of missing links are different to new links. In this work, we propose a new method to improve the performance of topology methods with time-aware graph. This method considers not only the current graph topology but also the historical changes of the graph. To make use of the changes of the graph, our solution first extract those important nodes based on a concept "active node". Then, we use this "active" information to create time-aware graph. Experimental result shows that our method could increase the performance of traditional algorithms.