Transfer learning across graphs drawn from different distributions (domains) is in great demand across many applications, yet the empirical performances vary and the in-depth understanding has been lacking. In this talk, I will first introduce our recent efforts on using inductive graph neural networks (GNNs) to solve the general cold-start problem of isolated "tail" nodes, transferring knowledge from the "head" nodes in the same graph yet with much richer neighborhood information. I will then dive into a special structured case of of graph transfer learning: given two graphs where nodes in the first graph (source) are well connected to each other while only sparse links are observed in the second graph (target), our aim is to predict links in the target graph by leveraging the source graph's richer information. We demonstrate that selectively leveraging the structural overlap will yield consistently stronger performance over a few common alternatives. Those techniques have been validated on both public academic benchmarks and real e-commerce datasets.