The rise of the Internet has ushered in an era of unprecedented data proliferation, presenting users with a daunting information overload dilemma when navigating a vast array of choices. This phenomenon has underscored the need for effective recommendation systems to assist users in navigating this abundance of information. Indeed, recommender systems have emerged as indispensable technologies for diverse online platforms, tasked with predicting user engagement with various projects. Many recommendation algorithms based on graph neural networks predominantly rely on user and item ID information to extract features, overlooking valuable user-item features embedded in additional data such as review texts. Alternatively, some approaches utilize auxiliary information to augment network training, adding complexity to the process. While models like GCMC and those presented in literature attempt to aggregate synergistic signals at different rating levels separately, they still do not adequately address the discrepancies in user viewpoints when amalgamating signals from multiple levels. Building upon the aforementioned considerations, this paper introduces an interest-aware message passing recommendation model. The proposed model employs two distinct methods to partition the graph into subgraphs: one for partitioning users and another for partitioning items separately. Through high-order graph convolution operations conducted within these subgraphs, the model aims to learn node representations for users and items, leveraging both weak user-item relationships within subgraphs and weak item-item relationships between subgraphs, thereby enhancing the expression of users' interests and preferences.