Graph Relearn Network: Reducing performance variance and improving prediction accuracy of graph neural networks

Zhenhua Huang, Kunhao Li, Yihang Jiang, Zhaohong Jia, Linyuan Lv, Yunjie Ma

Abstract: Highlights•Identified a new cause that hinders the efficiency and instability of current graph neural networks.•Proposed a graph relearn network that alleviates the oscillation of some nodes’spredicted classes during GNN training.•Found the unstable nodes are frequently located at the peripheries of clusters, and at the junctions between different clusters or communities.•Significantly improve the performance stability and accuracy of advanced GNNs and achieve new SOTA performances.•Extensive experiments are conducted on various real-world datasets.