Ride-hailing service is becoming a leading part in urban transportation. To improve the efficiency of ride-hailing service, accurate prediction of transportation demand is a fundamental challenge. In this paper, we tackle this problem from both aspects of network structure and data-set formulation. For network design, we propose a spatial-temporal attention multi-graph convolution network (STA-MGCN). A spatial-temporal layer in STA-MGCN is developed to capture the temporal correlations by temporal attention mechanism and temporal gate convolution, and the spatial correlations by multigraph convolution. A feature cluster layer is introduced to learn latent regional functions and to reduce the computation burden. For the data-set formulation, we develop a novel approach which considers the transportation feature of periodicity with offset. Instead of only using history data during the same time period, the history order demand in forward and backward neighboring time periods from yesterday and last week are also included. Extensive experiments on the three real-world datasets of New-York, Chicago and Chengdu show that the proposed algorithm achieves the state-of-the-art performance for ride-hailing demand prediction.