<p>Graph Neural Networks (GNNs) is a promising learning paradigm targeting to learning on graphs, by aggregating additional neighboring messages to supplement each node to generate expressive node representations. Despite achieved great success and wide applications in various graph based tasks, GNNs suffer from the problem of lacking of supplementary information due to the insufficient amount of neighbors, especially for neighbor starving cases in sparse graph. Even conventional graph augmentation methods can enhance the graph diversities, most approaches based on dropping nodes or edges intrinsically aggravate the challenge of insufficient neighbors. Moreover, existing graph data augmentation techniques suffer from semantics and label changing problems since they directly manipulating the original graph. To address above issues, we propose a novel Virtual Graph Augmentation method for Graph Neural Networks, named VGA-GNNs. Specifically, we train a conditional generative model to generate a parallel virtual graph for the original graph. Furthermore, a virtual topology augmentation strategy is further conducted on virtual graph to create a augmented virtual graph and enhance the structural diversity of virtual graph, which can preserve the integrity and invariability of the original graph when performing augmentation manipulation. In the training stage of GNNs, we are able to enlarge neighborhood and provide additionally supplementary information for nodes in original graph by building directed connection and message passing from the augmented virtual graph to the original graph. Extensive experiments on benchmark graph datasets demonstrate that the proposed method achieves significant performance improvements when various GNNs are used as the backbones.</p>