The deployment of high-performing recommender systems at scale is hampered by the computational complexity of traditional models and their inability to adapt rapidly to evolving user preferences. This work introduces an advanced hybrid framework designed to deliver highly personalized and efficient recommendations by leveraging synergistic computational techniques. The framework commences with Louvain Community Detection to rapidly and effectively partition the massive user-item interaction graph into cohesive, preference-homogeneous user communities, substantially reducing the search space for neighborhood analysis. Within these smaller sub-graphs, a novel Improved Neighborhood Heterogeneity-Sensitive Metric (INHSM) Similarity corrects for user and item popularity biases, yielding superior similarity estimates. Simultaneously, a global, Stochastic Gradient Descent (SGD)optimized Matrix Factorization model captures underlying latent factors. The final prediction is generated by a dynamic fusion of local INHSM neighborhood predictions with global SGD-based latent factor predictions, achieving state-of-the-art performance with superior accuracy, improved computational efficiency, and enhanced recommendation diversity.