Invariant Graph Neural Network for Out-of-Distribution Nodes

Zhengyu Chen, Yishu Gong, Liangliang Yang, Jianyu Zhang, Wei Zhang, Sihong He, Xusheng Zhang

Abstract: GNNs are effective for semi-supervised learning tasks on graphs, but they can suffer from bias due to distribution shifts between training and testing node distributions. In this paper, we propose the Invariant Graph Neural Network (IGNN) to address the issue of bias in GNNs. Specifically, IGNN learns the correlation of invariant features in different environments, where the spurious correlation changes in different environments. IGNN contains two components: the invariant graph partition component learns different graph environments and the invariant graph learning component regularizes the graph neural network to learn invariant graph representation in these environments. Extensive experiments have shown that the IGNN outperforms other methods for out-of-distribution nodes on several benchmark datasets.