The cold-start problem caused by data sparsity remains a significant challenge in recommendation systems. In this paper, we propose DiGcl-H, a data integration approach based on graph contrastive learning, designed to enhance recommendations for sparse groups by addressing both item cold-start and user cold-start challenges. Our integration approach innovatively captures the latent individual-group correlations in heterogeneous social networks by leveraging the user-item bipartite graph and the user social network graph. Additionally, it refines the cold item/user fusion optimization strategy through contrastive learning, effectively improving the integration of cold items and users within the same group without adversely affecting other group recommendations. Extensive experiment results on two real-world social recommender benchmark datasets, Filmtrust and Epinions, demonstrate that DiGcl-H is competitive with other state-of-the-art methods in cold-start scenarios, with Hit Ratio@50 increased by around 20%.