With the rapid development of the Internet, the computational analysis of social networks has grown to be a salient issue. Various research analyses social network topics, and a considerable amount of attention has been devoted to the issue of link prediction. Link prediction aims to predict the interactions that might occur between two entities in the network. To this aim, this study proposed a novel path and node combined approach and constructed a methodology for measuring node similarities. The method was illustrated with five real datasets obtained from different types of social networks. An extensive comparison of the proposed method against existing link prediction algorithms was performed to demonstrate that the path and node combined approach achieved much higher mean average precision (MAP) and area under the curve (AUC) values than those that only consider common nodes (e.g. Common Neighbours and Adamic/Adar) or paths (e.g. Random Walk with Restart and FriendLink). The results imply that two nodes are more likely to establish a link if they have more common neighbours of lower degrees. The weight of the path connecting two nodes is inversely proportional to the product of degrees of nodes on the pathway. The combination of node and topological features can substantially improve the performance of similarity-based link prediction, compared with node-dependent and path-dependent approaches. The experiments also demonstrate that the path-dependent approaches outperform the node-dependent appraoches. This indicates that topological features of networks may contribute more to improving performance than node features.