Social recommendation is an effective method to improve recommendation accuracy and recommendation system performance in recommender systems, attempting to combine user-item interactions with social links to reduce data sparsity and cold-start problems. Previous research approaches on social recommendation model the fusion of social information and user-item interactions, however, they ignore the problem of inconsistent social relationships, which affects the accuracy of recommendations. To consider the consistency of user social relationships, this paper proposes a graph neural network-based social recommendation model for user homogeneity, which obtains consistent embeddings of users at the contextual and relational levels through graph neural networks and relational attention. In social recommendation, through experiments on two mainstream datasets, we can demonstrate that our model outperforms the comparison model.