<p>Multimodal recommendation systems leverage heterogeneous data sources—such as textual descriptions and visual content—to alleviate cold-start and data sparsity challenges inherent in collaborative filtering. Existing approaches, however, tend to overemphasize local user-item interactions, thereby limiting their capacity to capture semantic content and global structural dependencies. We aim to capture both local interactions and high-order global dependencies among users and items.In this paper, we introduce CGTC (Collaborative Global hypergraph sampling with Type-aware sampling and Contrastive learning), a unified framework that integrates behavioral signals, collaborative patterns, and semantic relationships through multi-graph fusion and global hypergraph sampling. Specifically, CGTC constructs three complementary graphs–user–item interaction, item semantic similarity, and user co-occurrence–fused into a unified hypergraph. Item semantic graphs are computed per modality via cosine similarity and integrated for cross-modal relations, while the user co-occurrence graph retains only the top-k neighbors per user based on shared interactions. A type-aware global sampling strategy leverages hyperedge types to guide attention and aggregation, enabling the discovery of heterogeneous high-order relations. Additionally, an intra-modal contrastive learning mechanism is employed to enhance representation consistency within each modality. Extensive experiments on three public Amazon datasets—Clothing, Baby, and Sports—demonstrate that CGTC consistently surpasses state-of-the-art baselines, achieving up to 16.24% improvement in Recall@20 and 13.33% in NDCG@20 on the Clothing dataset. These results demonstrate the effectiveness of our model.</p>