FL-GNNs: Robust Network Representation via Feature Learning Guided Graph Neural Networks
Beibei Wang, Bo Jiang, Chris Ding
Abstract
Graph Neural Networks (GNNs) have been widely developed and grown rapidly to address representation and learning for attribute graph data <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$G(A,X)$</tex-math></inline-formula> . However, existing studies on GNNs mainly focus on the message passing on graph <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$A$</tex-math></inline-formula> for layer-wise propagation while pay less attention to the robust learning for the input features <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$X$</tex-math></inline-formula> , which thus make existing GNNs often perform susceptibility w.r.t feature noises and adversarial perturbations in <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$X$</tex-math></inline-formula> . In this article, we propose a novel Feature Learning guided Graph Neural Networks (FL-GNNs) by incorporating robust feature learning into GNNs. The core of FL-GNNs is trying to recover (or learn) a more clean and optimal feature data <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$Z$</tex-math></inline-formula> from input features <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><tex-math notation="LaTeX">$X$</tex-math></inline-formula> that better serves GNNs learning by jointly conducting feature reconstruction and GNNs' learning simultaneously. FL-GNNs is general and can be incorporated into any specific GNN models to enhance their robustness. An efficient algorithm has been derived to optimize FL-GNNs. Experimental results show that FL-GNNs can obviously enhance the robustness of existing GCN and GAT w.r.t feature noises and adversarial perturbations.