Graphs depict how different entities connect and interact with one another, and enable crucial structure-based prediction tasks, including node classification (NC), link prediction (LP), and community detection (CD). With the blooming and advances of deep learning, novel Graph Neural Network (GNNs) models, which learn the representations of nodes, are invented and widely applied on social networks. How powerful are GNNs on data science? In this talk, I will utilize our recent research outcomes to exhibit what, where, and how GNNs can benefit a variety of tasks on various data science tasks. First, we will demonstrate that hierarchical semantics-preserving GNNs are able to significantly boost the performance of typical NC, LP, and CD tasks. Second, we will show that GNNs can be applied to better model and exploit diverse relationships between various types of nodes in the realms of recommender systems and financial technology. Third, through the tasks of fake news detection, cyberbullying detection, and air quality forecasting, we will further exhibit that GNNs are still powerful even when the graphs cannot be observed. At the end of this talk, I will briefly point out the future directions of GNN-empowered data science.