Accurate prediction of user traffic in cellular networks is crucial to improve the system performance in terms of energy efficiency and resource utilization. However, existing work mainly considers the temporal traffic correlation within each cell while neglecting the spatial correlation across neighboring cells. In this letter, machine learning models that jointly explore the spatio-temporal correlations are proposed. Specifically, several recurrent neural network structures are utilized. Furthermore, a multi-task learning approach is adopted to explore the commonalities and differences across cells in improving the prediction performance. Base on real data, we demonstrate the benefits of joint learning over spatial and temporal dimensions.