We focus on predicting demands of bicycle usage in Velib system of Paris, which is a large-scale bicycle sharing service covering the whole Paris and its near suburbs. In this system, bicycle demand of each station usually correlates with historical Velib usage records at both spatial and temporal scale. The spatio-temporal correlation acts as an important factor affecting bicycle demands in the system. Thus it is a necessary information source for predicting bicycle demand of each station accurately. To investigate the spatio-temporal correlation pattern and integrate it into prediction, we propose a spatio-temporal network filtering process to achieve the prediction goal. The linkage structure of the network encodes the underlying correlation information. We utilize a sparsity regularized negative binomial regression based variable selection method to learn the network structure automatically from the Velib usage data, which is designed to highlight important spatio-temporal correlation between historical bicycle usage records and the bicycle demands of each station.Once we identify the network structure,a prediction model fit well with our goal is obtained directly. To verify the validity of the proposed method, we test it on a large-scale record set of Velib usage.