To facilitate big data analysis, we need both advanced computational methods to address scalability; as well as new statistical models to extract hidden structures from the data. In this talk, I present two research threads to address challenges from both computational and statistical aspects in modern data analysis: (1) An uniformly-optimal stochastic first-order method for large-scale online prediction and its implementation in a distributed environment. (2) A computationally efficient method for predicting dynamic graphical models from complex data. I will also talk about the applications of the proposed methods, such as text mining and climate data analysis.