Abstract: Path-based and representation-based reasoning are two major methods on knowledge inference. A combination of both algorithms can improve the accuracy of knowledge reasoning. However, there are still some problems, such as inefficiencies in learning, low prediction accuracy and over-fitting of the model. A representation learning method based on path selection was proposed in this paper to further filter the path feature information, to hold the key paths and to use the balance parameter to process the triples of missing path information. In this paper, a public data set was used to test the model. Experiments show that the model can effectively improve the generalization ability and accuracy.