Graph neural network based on graph kernel: A survey

Graph data are pervasive in real-world scenarios, and research on graph data has become a research hotspot. Over the past few decades, significant advancements have been made in the graph domain, particularly in the development of graph kernels and graph neural networks. But they also face challenges, such as graph kernel is difficult to learn complex interactions, and too many parameters of graph neural network lead to poor optimization, etc. Therefore, integrating them has become an important strategy. Existing reviews in the published literature primarily concentrate on either graph kernels or graph neural networks individually, with no mention of the graph neural network methods based on graph kernels. This paper starts from the basic knowledge, presents the challenges they encounter, and analyzes the existence of complementary perspectives between them, thus confirming the feasibility of the integration strategy. Following this, this paper organizes some important methods of graph neural networks based on graph kernels in recent years in terms of expressiveness, performance, and applications. In addition we have substantiated the actual effectiveness with experimental results. Lastly, we explore future research directions. We have also collected papers and open source code resources on graph kernel based graph neural network methods in recent years at https://github.com/bigdata-graph/GNN_GK.