A study on K-hop structural similarity to alleviate over-smoothing in GNN

Abstract

Oversmoothing has been pointed out as one of the causes to exacerbate the degradation of the classification accuracy of GNNs. To mitigate this problem, many efforts have been made. For example, GraphSNN uses subgraph similarity as weights during aggregation in the message-passing framework, and KP-GNN makes K-hop GNN more expressable using peripheral subgraphs. In this paper, we propose a way to calculate the similarity of the $k$-hop nodes in order to extend the idea of GraphSNN to K-hopGNN and reduce the negative effect of Oversmoothing even more. We calculate similarity using the list of degrees of the neighboring nodes to consider the structural information of the graph. As a result of graph classification experiments, we found that our method can improve graph classification accuracy and reduce the effects of Oversmoothing.