An Improved Transformer Model Based Graph Node Classification Method

In recent years, Transformer models have been widely developed in various fields, such as natural language processing (NLP) and computer vision (CV). However, its application in graph data structures is still limited. It is crucial to develop an improved Transformer model for graph data. Considering the weakness that the original Graph Transformer (GT) is difficult to capture location information of nodes significantly, this paper combines node positional embedding in each layer's output. Then, in order to select more efficient neighbor nodes, a mask matrix is introduced to reconstruct attention score matrix. Finally, the group normalization is applied to provide faster training speed and better generalization performance. A number of experiments on two reference graph datasets demonstrate that the proposed GT model has better performance compared to other standard GNN s and original G T.