<p>With the development of graph learning technology, different modalities of content information such as visual and textual can be learned. Applying graph learning techniques to recommendation has become a research trend. To address the challenges of data sparsity and multimodal data fusion in recommendation, we propose a novel graph collaborative filtering framework based on multi-view information fusion. This framework utilizes user social view to mine shared preference features between users. At the same time, based on extracted text and image features, it constructs textual and image semantic graph by utilizing the similarity of features within each modality, achieving effective fusion of multi-view semantic information and comprehensively mining multi-view and multimodal information. Then, we designed a contrastive learning mechanism to strengthen the features between different views, effectively improving the distinguishability between semantic features of different views. This framework provides a scalable solution for recommender systems in sparse data scenarios, where it can improve recommendation accuracy while reducing cold-start bias. For data-sparse subsets, our model outperforms baselines more significantly on sparse subsets (14.8–34.2% relative improvement on HR@10) than on the full dataset (2.3–7.2% on HR@10). Extensive experiments on distinct real-world datasets have provided solid evidence to confirm the effectiveness of our model.</p>