In order to solve the problem that the link prediction method based on single relational path cannot mine the influence of different paths in the knowledge map, a link prediction method based on multi relational path is proposed. Firstly, the similarity index based on path information is used to calculate the similarity between all relational paths. Then, the relationship projection between different paths is extended to the new path projection and path constraints, and the training process is performed by using random gradient descent, so that the explicit features between different paths can be screened out through low dimensional representation learning in implicit space. The validation analysis is carried out on Enron email data set and National Natural Science Foundation data set. Experimental results show that, compared with other path link prediction algorithms, the maximum improvement of map and AUC is about 20%, showing higher prediction accuracy.