| This thesis focuses on Data-Driven Approach in Networking (DDN), a new paradigm for network control and management. DDN is to use network measurement and user behavior data, based on machine learning techniques, and control/optimization mechanisms, to solve network control and management challenges. Modern networks are extensively monitored, providing us with a great amount of networking data. The emergence of software defined networks and cognitive radio increases the flexibility of network control. With these advancements, the idea of using learning techniques to intelligently control the network has become attractive. Applying learning and control techniques to network data, DDN tackles traditional networking problems from a new perspective. Existing work has proved that DDN could produce efficient algorithms for complex systems, generate adaptive policies for changing requirements, and avoid expensive network measurements.;This dissertation first presents an overview of the DDN paradigm, including its components and applications in the literature. Data collection, learning techniques, and control algorithms are the three major components of DDN. We present three common applications of DDN, and discuss related existing work. Then, we present two concrete examples, UPDATE and Earlybird, to demonstrate how to apply DDN to real-life networking problems. First, UPDATE efficiently schedules network transfer by learning from history user network profiles. Second, Earlybird learns the users' social content preferences and mobile network profile to prefetch embedded content in mobile social applications. |
