We consider the problem of data augmentation for graph neural networks. While other areas of deep learning, especially computer vision, has benefited greatly from data augmentation, graph deep learning is limited by the complex structure and nodes dependency of graph datasets. The works studying data augmentation for graphs focus on ways to train the model on generated graphs or disturb the original one by masking attributes or dropping nodes/edges. Our work studies data augmentation for graphs in the latent space. We propose an architecture to train a graph neural network including a data augmentation component which also has a denoising and structure enhancer function. Our architecture leverages the supervised feedback to introduce the graph neural network component to a new latent representation of the graph with lesser irrelevant connections on the graph domain. We perform experiments on datasets from several domains with different sizes and show that our architecture improves the performance over a variety of plain graph neural networks on node classification.