**4. Experiments**

To evaluate the performance of our framework, we conduct experiments on a realworld dataset from New York Citi Bike. Details of multi-source open data are in Table 4. Bike station data are collected from June 2013 to November 2018, and stations operating for less than six months, or with a monthly average demand of less than 300, are removed. Batches can be realized as the time period of a relatively large number of bike stations construction. Stations with established dates from June 2013 to July 2015 are the origin. From Batch 1 to 3 prediction, we divide stations in the training set and testing set according to their established date. For instance, in Batch 2 prediction, stations established earlier than August 2016 are training data, and the other stations established during August 2016 to September 2016 are testing data. We retrieve multi-source open data from Citi Bike (the

bike-sharing system in New York), bike routes data, and Facebook Place API. Detailed datasets are listed in Table 4. The settings for radius *r* of the reachable station region are 500 m, and we extract the top-15 nearby station features in our experiment.
