Graph Neural Networks (GNNs) have shown great potential in social recommendation systems by integrating node information and topological structure. This paper proposes a novel GNN framework called GraphRec to address three challenges in social recommendation: (1) combining user-item and user-user social graphs, (2) jointly capturing interactions and opinions in the user-item graph, and (3) handling heterogeneous strengths of social relations. GraphRec jointly models these two graphs and their heterogeneous strengths, and experiments on two real-world datasets demonstrate its effectiveness. The framework uses attention mechanisms to capture interactions and opinions, and differentiates social relations based on their strengths. The model is trained using a combination of user and item modeling components, and the results show that GraphRec outperforms existing baselines in rating prediction. The paper also discusses the impact of model components and hyperparameters on performance, and highlights the importance of social network information and attention mechanisms in improving recommendation accuracy. The proposed framework provides a principled approach to social recommendation and demonstrates the power of GNNs in learning from graph data.Graph Neural Networks (GNNs) have shown great potential in social recommendation systems by integrating node information and topological structure. This paper proposes a novel GNN framework called GraphRec to address three challenges in social recommendation: (1) combining user-item and user-user social graphs, (2) jointly capturing interactions and opinions in the user-item graph, and (3) handling heterogeneous strengths of social relations. GraphRec jointly models these two graphs and their heterogeneous strengths, and experiments on two real-world datasets demonstrate its effectiveness. The framework uses attention mechanisms to capture interactions and opinions, and differentiates social relations based on their strengths. The model is trained using a combination of user and item modeling components, and the results show that GraphRec outperforms existing baselines in rating prediction. The paper also discusses the impact of model components and hyperparameters on performance, and highlights the importance of social network information and attention mechanisms in improving recommendation accuracy. The proposed framework provides a principled approach to social recommendation and demonstrates the power of GNNs in learning from graph data.