Graph collaborative filtering has achieved significant success in improving the performance of recommender systems by performing message passing recursively along the user-item interaction edges to capture the user’s preferences for items to improve recommendation performance. However, in actual application, user behavior data is usually sparse and noisy, which limits the accuracy of recommendations. To address these issues, Contrastive Learning (CL) has been used in recommendation. However, most CL-based approaches use heuristic data augmentation to construct contrast pairs, that may result in the loss of important information and lacking the effective utilization of node higher order relationships, which have demonstrated significant usefulness in recommendation tasks. In this paper, we propose a novel method, graph Generation and Multi-graph fusion Contrastive Learning for recommendation, called GMCL. The method learns collaborative relationships among users by constructing a user-user graph, and uses an adjustable strategy to combine embeddings learned from user-item interaction graph with collaborative relationships. To mitigate the over-correlation problem caused by merging collaborative relationships, we perform feature de-correlation on the fused embeddings. In addition, to build useful CL tasks, the method proposes a novel adaptive generation-fusion contrast learning paradigm. Experimental results reveal that GMCL is valid and surpasses current collaborative filtering approach significantly.