Node classification based on structure migration and graph attention convolutional crossover network

Due to the sparse structure of graph and GCN (Graph Convolutional Networks) does not consider neighbor node specificity, graph nodes are over-smoothed after passing through the GCN network, especially when the number of network layers increases. Meanwhile GCN cannot propagate over a wide range and capture long-range information. To solve the above problems, this paper proposes a node classification algorithm based on Structure Migration and Graph Attention CoNvolutional crossover network (SM_GACN). Firstly, a Pearson-adjacent structure migration method based on attribute correlation and first-hop neighborhood sparsity is proposed. By measuring Pearson coefficient between the attribute vectors of unlabeled and labeled nodes, the most relevant labeled nodes with unlabeled nodes are selected. And then by comparing the sparsity of the first-hop neighborhoods of the two nodes, this method determines whether or not to perform structure migration on this unlabeled node. Secondly, a graph attention convolutional crossover network based on graph attention and graph convolution is designed. This network extracts features by alternating graph attention and graph convolution layer, and then completes one linear mapping by graph convolution layer. Finally, a combine-training method based on inductive learning is constructed. The original structure graph is combine-trained with the migrated structure graph, and the GCN with the crossover network, which incorporates inductive learning. Comparing with eleven algorithms on four datasets, the experimental results show that SM_GACN has higher classification performance on the graph node classification task.