Existing collaborative filtering models based on the graph neural network (GNN) suffer from noise and data sparsity in the original graph. A novel graph neural network recommendation model (GICL) was proposed, employing singular value decomposition to retain the principal collaborative relationships from global information and suppress local noise. To capture latent relationships between nodes, representative embeddings were learned from both the graph structure and semantic space as augmented neighbors, and contrastive learning tasks were constructed based on these expanded neighbors. Specifically, a structural contrastive view was created by aggregating homogeneous node information using GNN, while a semantic contrastive view was constructed by partitioning the original graph into semantic subgraphs and using their prototypes. The graph collaborative filtering task was jointly trained with two contrastive auxiliary tasks to alleviate the data sparsity issue. Experimental results on four public datasets demonstrate that the GICL consistently outperforms many mainstream recommendation models.