Traffic congestion is a major problem throughout the world. It impedes our mobility, pollutes the air, wastes fuel, and hampers economic growth. Thanks to the dramatic rise of sensors like loop detectors, mobile phones, traffic cameras, and self-driving cars, for the first time, a flood of truly real-time data can help eliminate our traffic congestion problems. In our research center at USC, for the past five years, we have been collecting data from more than 15,000 loop detectors installed on the highways and arterial streets of Los Angeles County. Each time a car passes through the loop detector, information such as traffic speed is collected and stored in databases. However, there are many challenges in collecting, cleaning, storing, and learning from such a large-scale dataset. To address these challenges, we have developed an archived data management system. This system is designed to handle high fidelity data. In this paper, we studied the traffic prediction problem. Our objective is to predict the traffic situation in the near future for each road segment. To achieve this, we propose a latent space model that learns the spatial and temporal attributes of vertices in the road network. These attributes can estimate how traffic patterns vary. With the success of our prediction technique, we can utilize it for various applications such as better route navigation, urban planning, and traffic regulations.