Learning accurate 3D shapes from sparse and incomplete point clouds is challenging and meaningful, on account that the point clouds with low resolution always lack representative and informative details. This paper presents a novel deep auto-encoder called TGNet, which is formulated based on a tree-based generative adversarial network (GAN), to address self-supervised learning tasks on the point cloud with low sparsity. On the encoder side, we employ a PointNet-based framework to intensively capture the global representations. To better infer the spatial information in latent space, we propose a spectral graph learning module in with due consideration to graph topology. Further, we present a new loss that combines Wasserstein metric and multi-resolution Chamfer distance to better estimate global 3D geometry and structural details. The proposed TGNet achieves state-of-the-art performance for various point cloud learning tasks. Qualitative and quantitative evaluations demonstrate the novelty of the proposed model.