Traditional Graph Neural Network (GNN)-based recommendation systems encounter significant challenges in managing high computational costs and scalability when modeling complex, heterogeneous user-item interactions. These models frequently fall short in capturing the diversity of relationship types and multi-level contextual information, resulting in limited performance gains. To overcome these limitations, we present the Heterogeneous Light Graph Convolutional Network (He-LiGCN), a streamlined framework designed to enhance both accuracy and scalability. Leveraging multi-level graph coarsening and sophisticated contextual embeddings, He-LiGCN efficiently models intricate user-item dependencies with substantially reduced computational demands. Experimental results demonstrate a 6.5% improvement in Recall@10 and an 8.9% increase in NDCG@10 over baseline models, along with a reduction in training time by up to 20%, underscoring its effectiveness and efficiency in large-scale recommendation scenarios. By advancing the efficiency and depth of graph-based recommendations, He-LiGCN represents a practical and scalable solution for the next generation of personalized recommendation systems.