A bike sharing system deploys bicycles at many open docking stations and makes them available to the public for shared use. These bikes can be checked-in and checked-out at any of the docking stations. Predicting daily visits is important for service providers to optimize bike allocation and station maintenance. In this paper, we formulate this prediction problem as a regression task. Through data analysis, we develop several features that are very helpful in predictions. Moreover, we demonstrate that there are significant differences among the patterns of visits at different stations. To improve prediction accuracy, we propose station-centric augmented with global feature transformation. The gradient boosting decision tree (GBDT) and neural network (NN) techniques are leveraged to extract global features. The experimental results demonstrate that the proposed model offers better prediction performance compared to two baseline approaches.