With the popularity and development of the internet and mobile terminals, people can access a lot of information through them every day. Recommender systems have become one of the important technologies for various online platforms that aim to predict whether a user will interact with an item or not. Among them, collaborative filtering-based models have made effective progress in learning user and item representations by modeling historical user-item interactions. Recently, models based on GCN have been effective in recommendation, and the main function of GCN models is to improve the embedding representation of users and items by iteratively aggregating feature information from neighbors using graph connectivity to extract additional information. However, in previous works, dividing users into subgraphs without intersection only leads to a partial loss of information, ignoring the potential connections that may exist between different groups of users; and because only users are divided, the influence of commodity factors on the purchase outcome at the time of purchase is ignored in the learning process. Based on the above considerations, in this paper we propose a message-passing recommendation model. The model uses intersects users and items in separate subgraphs and uses an optimized attention mechanism to obtain the final node embedding to optimize the embedding representation by introducing multiple embedding propagation layers that encode higher-order connectivity relationships. We conduct extensive experiments to evaluate the proposed model. The results show that our model can effectively improve the performance of the recommendation.