GraphMix: Improved Training of GNNs for Semi-Supervised Learning

We present GraphMix, a regularization method for Graph Neural Network based semi-supervised object classification, whereby we propose to train fully-connected network jointly with the graph neural via parameter sharing and interpolation-based regularization. Further, provide theoretical analysis of how GraphMix improves generalization bounds underlying network, without making any assumptions about "aggregation" layer or depth networks. experimentally validate this by applying various architectures such as Convolutional Networks, Attention Networks Graph-U-Net. Despite its simplicity, demonstrate that can consistently improve closely match state-of-the-art performance using even simpler across three established benchmarks: Cora, Citeseer Pubmed citation datasets, well newly proposed datasets: Cora-Full, Co-author-CS Co-author-Physics.