To address the problem that existing graph neural network methods only focus on the spatial or spectral domain when extracting features, ignoring the advantages of combining the two, resulting in less comprehensive feature representations. In this paper, we present a graph neural network incorporating deep local topology (IDLT-GNN), which aims to exploit the more complex and richer connections among individual nodes in graph data. Specifically, IDLT-GNN firstly introduces Variational graph auto-encoders (VGAE) to learn the node features of the existing graph data from the null domain, which helps to better capture the implicit structure and feature relationships among nodes. Secondly, the Eigen-GNN method is introduced to extend the initial basis from the spectral domain by directly connecting the abstract eigenvectors of the graph structure matrix with the node features to enhance the ability to maintain the graph structure, which helps to capture the complex relationships among the nodes and enhances the model generation capability. The resulting graph structure data is then re-learned by graph convolutional neural network and multilayer perceptron to obtain richer node features. Finally, experiments on a publicly available benchmark dataset demonstrate that the method improves 1.5 and 2.5% in the node classification task compared to the traditional method, proving the superiority of IDLT-GNN in learning the null and spectral domain features of graph data.