Multi-view graph imputation network

Graph data in the real world is often accompanied by the problem of missing attributes. Recently, self-supervised graph representation learning, implementing data imputation according to observable nodes, has become a paradigm for studying attribute-missing graphs. However, existing methods directly encode attribute-missing graphs and then impute attribute-missing nodes in the latent space, which often increases the uncertainty of node embedding information, thereby limiting the performance of attribute imputation. To address this issue, we propose a novel method named Multi-view GrAph impuTation nEtwork (MATE), which performs attribute imputation in the input space for attribute-missing graphs. Specifically, we first employ parameter initialization and graph diffusion in the input space to generate relatively complete multi-view from both the attribute and structure levels. To provide reliable guidance in each epoch for parameter initialization, we propose a Dual Constraint Strategy (DCS) that maximizes the consistency of node embeddings between two views. In this way, the learning of parameter initialization and node embedding promotes each other, thus effectively improving the quality of attribute imputation. Extensive experiments on four benchmark datasets demonstrate that our proposed MATE achieves state-of-the-art performance. The corresponding code is available at https://github.com/XinPeng97/MATE.