Path Integration Enhanced Graph Attention Network

Hui Wang, Peng Zhou, Junbo Ma

Abstract: Graph attention networks are a deep learning method for processing graph data. By learning the relationships between neighbouring nodes in the graph, GATs have been widely used in many fields. However, the graph attention network has the problem of information lag in the process of information aggregation, which degrades the performance of the graph attention network. Referring to the ideas of Feynman path integral theory in physics, we proposed a new graph attention method called PaInGAT to solve the above issue by introducing a new neighbor information aggregation mechanism. Specifically, we improve the neighbour node aggregation mechanism of traditional graph attention networks by calculating the path integral from the source node to the target node to obtain the attention factor, and update the information of multi-order neighbours to the central node directly by the attention factor of the current state at each layer. Through experimental demonstration combining different downstream tasks, our method achieves excellent results on several datasets, demonstrating its effectiveness and advancement.