Abstract: The application of the knowledge graph to the recommendation algorithm can effectively reduce the problem of data sparsity and user cold start barriers in the recommendation system. This paper proposes a weighted recommendation method based on the neighborhood expansion feature. This method added the correlation strength between the relationship and the entity to complete the neighborhood representation of the items features in the knowledge graph convolutional network (KGCN). We combined it with the results of the model (RippleNet) that expanded users features using a weighted method. The KL (Kullback-Leibler) divergence was introduced as the deviation. The experiment was carried out on two datasets of MovieLens-1M and Book-Crossing. The results show that the proposed method is improved on AUC, F1 and ACC compared with the baseline model.