Understanding Distributions of Traffic and Mobility Data

Abstract
Data serves as a crucial foundation for research, particularly when it comes to analyzing complex systems. The objective of engaging with data is twofold. First, we aim to gather real-world observations to develop models.  Then, these models enable us to simulate and thus better comprehend the dynamics of the real world. This process of deriving insights from data essentially involves learning the underlying distribution of the collected samples (data). Seongjin Choi structured this presentation to explore two principal methodologies to achieve this understanding. In the first half of the presentation, he introduces one method called "Deep Probabilistic Forecasting," which aims to model network-level traffic state. Deep Probabilistic Forecasting involves assuming a probabilistic distribution and learning its parameters. In the second half of the presentation, Choi introduces another method called the "Deep Generative Model" (DGM), which aims to model agent-level trajectories. DGMs aim to train neural networks capable of generating synthetic samples that mirror the characteristics of the original data. Overall, the presentation aims to showcase the capabilities of both methodologies in capturing patterns and behaviors in transportation and mobility data.