Complex networks include biological information networks, collaboration networks and social networks. Studying the relationship prediction of complex networks helps predict relationship between proteins, find out cooperation relationship among scientists, as well as mine potential social networks. Currently, most relationship prediction algorithms are realized by similarity-based models, however, this type of algorithms based on network topology feature are explicitly constructed, which ignore latent information behind generated relationship. To solve this problem, we propose an enhanced naive Bayesian relation prediction model (ELNB), which defines a conditional probability to model the local sub-graph structure. It can effectively alleviate the independence assumption of LNB and realize a quantitative calculation of neighbors contribution. Experiments on artificial datasets and real datasets show that the proposed model is better than the baselines and some recently proposed models. Meanwhile，the idea of ELNB can be extended to other similarity algorithms based on common neighbor nodes, which provides a new method for the research of such kind of model.