Graph neural networks have been widely used in recommendation systems and have achieved outstanding performance. Graph convolutional neural networks (GCNs) are classic graph neural networks that can better extract features and representations from graph-structured data by stacking multiple layers of GCNs. However, stacking GCNs can lead to over-smoothing, making nodes difficult to distinguish and affecting model performance. Meanwhile, the natural noise in the user-item interaction graph also affects the model’s ability to learn better embeddings. In this paper, we propose the Layer Transformer-Powered Graph Convolutional Networks (LayerTrans) for recommendation to address the over-smoothing problem. The model dynamically uses cross-attention to extract information during the information propagation process of GCNs and then updates the nodes to learn better embedding representations of users and items. In addition, we use degree centrality to measure the importance of each node and calculate the probability of edge deletion between nodes based on their importance. This paper conducts sufficient experimental evaluation on four real public datasets, and the comparison results with multiple existing methods demonstrate the superior performance of our method.