LAUSR

Graph neural networks (GNN) have achieved remarkable success in various domains, yet incomplete node attribute data can significantly impair their performance. Graph completion learning (GCL) methods have been developed to address this issue, aiming to reconstruct missing node attributes based on existing structural relationships. However, the accuracy of these reconstructions is highly dependent on the quality of the initial graph structure, which often contains errors and inaccuracies. This paper proposes the evolving graph structure (EGS) framework for semi-supervised node classification with missing attributes. EGS dynamically reconstructs the attributes of the nodes and updates the graph structure through an alternating optimization approach. Specifically, we introduce a Dirichlet Energy function with dual constraints to formulate the objective function, which jointly optimizes node structure relationships and attribute reconstruction. Extensive experiments on five benchmark datasets, with different missing rates, and with seven GNN variants demonstrate the effectiveness of EGS, achieving state-of-the-art performance compared to existing GCL methods.