Understanding transportation network flow patterns with high-dimensional multi-source data

Nowadays, emerging technologies greatly improve mobility, but also make the transportation system more complicated than ever. Today’s systems commonly feature heterogeneity, interactions among multiple components, and high dimensionality. However, traditional transportation network modeling methods have three major limitations, making them less ideal for understanding transportation network flow patterns in today’s systems. First, traditional network models generally assume that recurrent traffic flows at the equilibrium follow one single dominating pattern, ignoring the heterogeneity in the data. Second, traditional models often focus on only one or two components of the whole urban system and are incapable of consuming multi-source data. Third, traditional traffic models mostly do not scale well with the size of the network, making them less applicable for high-dimensional cases. 

In correspondence to the three limitations, in this thesis we propose three novel methods for augmenting traditional transportation models so that they could better utilize the massive high-dimensional multi-source data of heterogeneous traffic flows. The first tool is a clustering algorithm in the context of transportation networks, which could be used to split a heterogeneous data set into homogeneous groups, allowing each to be modeled individually with traditional models. The second tool is a framework to laterally combine transportation network models and vehicle emission models to create an integrated model that is able to consume data sets from the whole urban system and produce insights on improving mobility and controlling environmental impacts at the same time. The third tool is a method to approximate path-based network models and hence to reduce the computation complexities of those models. With this method, traditional network models could scale well with the dimensionality of the network path flows, and would be more applicable for modeling large-scale networks. 

Data-driven transportation network modeling with high-dimensional multi-source data is an important and broad topic, on which there are a great amount of on-going efforts by researchers and practitioners. As such, we do not aim to propose a one-stop framework for this modeling problem. Instead, we make the aforementioned three tools reusable, composable, and widely applicable, so that they could enrich our modeling toolbox for a wide spectrum of modeling needs in the real world. In this dissertation, we examine the proposed methods on several real-world transportation networks, and demonstrate the performance, effectiveness, and applicability of those methods in practice.