In complex networks, identifying vital nodes is important to improve the reliability of the network and ensure its safe and effective operation. However, traditional centrality methods are one-sided and inaccurate, and centrality methods work differently in different networks. At once, many nodes have the same centrality value, making it difficult to distinguish differences between nodes. In order to better identify vital nodes and improve node resolution, inspired by graph structure learning, graph neural network was adopted to encode nodes and fit node influence, the propagation field learner was used to compute the propagation field of network nodes, and the degrees of node were observed as labels to train the model. Finally, the SI (susceptibility-infection) model was used for propagation simulation, and kendall's correlation coefficient was used as a measure of node ordering monotonicity. The results in real-world networks show that the algorithm has better better performance in measuring the sorting performance with the SI propagation model and the monotonicity of the sorting results with the kendall correlation coefficient, which improves the accuracy of vital nodes identification and the robustness of network node sorting.