Recommendation systems (RS) often rely on global models such as matrix factorization (MF) or graph neural networks (GNN) to capture user-item interactions. However, these approaches may struggle with data sparsity, cold-start users, and limited interpretability. In this study, we propose a novel framework that integrates spectral clustering with localized MF to improve recommendation accuracy and scalability. By constructing a user-user (or item-item) similarity graph and applying spectral clustering on its Laplacian representation, we identify latent communities within the data. Separate MF models are then trained within each cluster, capturing community-specific preferences. For user-item pairs across different clusters, a global fallback model ensures full coverage. We evaluate our method on multiple real-world datasets, including MovieLens, Yelp, Amazon, and Last.fm. Experimental results show that our approach consistently outperforms traditional MF, SpectralCF, NGCF, and LightGCN, particularly in sparse and cold-start settings. The proposed method also offers improved interpretability by revealing latent structures in the user-item space. This study demonstrates the effectiveness of community-aware modeling in enhancing recommendation performance using lightweight, interpretable techniques.