ViVA: Semi-supervised Visualization via Variational Autoencoders