Abstract: The existing knowledge inference based on knowledge graph embedding is generally to predict missing facts by embedding entities and relations into low-dimensional vectors. However, this method only uses zero-order logic to encode existing triples and cannot apply first-order logic rules. Although the integration of logical rules embedding can solve this problem, it also brings the problem of vulnerability of logical rules. Therefore, a joint embedding method of logical rules and inference paths is proposed in this paper. By jointly representing logical rules and paths in a unified embedding space, entities, relations, and learning logic rules are jointly represented from existing triples and horn rules. According to a given logical rule, we generate corresponding paths by using a walk strategy which integrates the confidence measure of the entity pair relationship path, omits intermediate entities, and obtains a highly correlated path set to supplement the semantic relationship between entity pairs. We also perform similarity calculations with the rule head relationship to obtain path associates embedding scores. In addition, during the optimization process, we also consider the confidence of each rule to ensure the usability of rule embedding. Finally, based on a large number of experimental results on four standard knowledge graph datasets, it was demonstrated that this method can effectively alleviate the vulnerability of logic and improve inference performance. Compared with suboptimal models, it achieves a maximum improvement of 2.0% in Hits@10 on the Kinship dataset and a maximum improvement of 2.2% in MRR on the UML dataset.