Sung-Eun Kim, Receptive Field Depth Selection in Graph Neural Network for Node Classification Task, Master's Thesis, Department of Industrial and Systems Engineering, KAIST, 2023 

Abstract

Graph neural networks (GNNs) have been shown good performance for semi supervised node classification task by aggregating neighbor information based on the graph structure. Since different nodes have different adjacent environments structurally and semantically, it is a problem that all nodes have the same receptive field. We observe that model prediction of specific nodes becomes unstable depending on which hop information is used. This is because specific nodes are in a adjacent environment that makes model prediction unstable, so adjusting the receptive field of these nodes helps improve overall performance. In this work, we analyze the components of adjacent environment that distabilize the classification of nodes, and propose post-processing of model prediction algorithm to correctly classify them. It shows good performance for unstable nodes through quantitative/qualitative comparison with existing studies.

@masterthesis{Kim:2023,

author = {Sung-Eun Kim},

advisor ={Il-Chul Moon},

title = {Receptive Field Depth Selection in Graph Neural Network for Node Classification Task},

school = {KAIST},

year = {2023}

}