Abstract: At present, the mainstream recommender system model needs to obtain enough data to have good performance. When the obtained data is sparse, the accuracy of recommendation results is poor. At the same time, better solutions are needed for pushing the projects newly added to the recommender system to potential users and obtaining the interests of new users. A recommender system model OCKG(recommender system based on object feature combination embedded and knowledge graph)is proposed.The model takes the users and the projects as the opposite object, embeds and obtains the correlation tags through the multi-dimensional information of the users and the projects respectively, and trains them to obtain the commonality of the same kind. At the same time, the embedded tags are weighted to spread different attribute features to the knowledge graph to increase the similarity strong model learning transfer ability, ranking the recommendation results horizontally and vertically according to the close correlation, so as to realize the prediction of recommendation results. Comparative experiments with two different public datasets demonstrate the validity of the proposed model under sparse data and cold start. The experimental results show that the model's recommendation performance and robustness are enhanced by reasonable feature combination and the propagation intensity on the knowledge graph.