Exploring attention mechanism for graph similarity learning

Graph similarity estimation is a challenging task due to the complex graph structure. Though important and well-studied, three key aspects are yet to be fully handled in a unified framework: (i) how to exploit the node embedding by leveraging both local spatial neighborhood information and the global context, (ii) how to effectively learn richer cross graph interactions from a pairwise node perspective and (iii) how to map the similarity matrix into a similarity score by exploiting the inherent structure in the similarity matrix. To solve these issues, we explore multiple attention mechanisms for graph similarity learning in this work. More specifically, we propose a unified graph similarity learning framework involving (i) a hybrid of graph convolution and graph self-attention for node embedding learning, (ii) a cross graph co-attention (GCA) module for graph interaction modeling, (iii) similarity-wise self-attention (SSA) module for graph similarity matrix alignment and (iv) graph similarity matrix learning for predicting the similarity scores. Extensive experimental results on three challenging benchmarks including LINUX, AIDS, and IMDBMulti demonstrate that the proposed NA-GSL performs favorably against state-of-the-art graph similarity estimation methods. The code is available at https://github.com/AlbertTan404/NA-GSL.