Recommendation systems chiefly rely on Collaborative Filtering techniques to depict user inclinations by analyzing historical interactions. This traditional approach has seen significant improvements with recent advancements that leverage Graph Neural Networks (GNNs). GNNs enhance CF by exploiting high-order interactions within user-item bipartite graphs [1]. However, existing GNN-based CF models face a critical limitation: the entanglement of depth and range, where increasing GNN layers simultaneously enlarges the receptive field, complicating the model and increasing computational demands. To address this, we propose a framework that decouples range from depth, using Personalized PageRank (PPR) to extract localized subgraphs with bounded ranges, allowing for deeper GNN layers without expanding the receptive field. Additionally, we introduce a learnable aggregation function that optimally integrates sum, max, and mean operations for improved message aggregation. Experiments on multiple datasets demonstrate our approach’s superior performance. The results indicate that our framework not only addresses the depth and range entanglement issue but also significantly boosts the effectiveness of GNN-based CF models.