Human mobility is a key for sustaining activities in modern societies. In terms of frequency and farness, it has been continuously growing in line with globalization, bringing enormous impacts on the evolution of human societies. Investigation in human mobility not only facilitates our understanding of human behaviors, but is also useful for studying epidemic spreading, conducting urban planning, etc.

While there are various activities contributing to human mobility, this thesis focuses on commuting, which accounts for the movement between the place of residence and the place of work. The commuting pattern reflects a population-level movement, which is relatively stable and regular. It, arguably, has the most significant impact on daily life as well as urban planning and, thus, attracts tremendous research interests. This thesis spans three aspects corresponding to the commuting pattern, namely, the modeling of the physical process behind the commuting behavior, the reconstruction of commuting networks, and the modeling of communication and consensus processes in the mobility-consequent pluralistic society.

A commuting network can be considered as an aggregated outcome of the population-level job region selection. Currently, there exist some useful models to describe and predict the commuting flow. However, self-loop flows are commonly excluded, due to either the limitation of the model or the difficulties in accurate prediction for this data, even though they contribute a high percentage in commuting. Moreover, while regional attractiveness is commonly considered as a basic and useful concept in commuting, it is roughly estimated by the population size, which in return, leads to poor performance in flow estimation.

To resolve these two fundamental issues, we propose an attractiveness-based mobility model to estimate commuting flows in all spatial ranges, empowered by a trip competition mechanism (TCM). The model includes attraction scores of regions in concern, obtained via optimizing the working population distributions. Its capability of capturing a variety of mobility patterns is verified by empirical data from three different countries, and its accuracy outperforms those of existing models. The quantified attractiveness is also found to be highly correlated with common socioeconomic indicators and is able to act as a distinct metric to characterize a region.

Then, combining the TCM concept with novel machine learning theories, we further develop an advanced commuting network reconstruction method. It utilizes a geographic competition graph (GCG) and a distance-tiered graph neural network (DtGNN). The GCG represents data in a graph, modeling the competition relationship behind the job selection process. The DtGNN is a novel design of GNN that utilizes distance information to realize the weights sharing and achieve node embedding for commuting flow prediction. The effectiveness of GCG and DtGNN is confirmed via extensive experiments on real-world data from the United States. Significant improvements are observed, as compared to both traditional commuting models and state-of-the-art machine learning based methods.

Undoubtedly, human mobility removes the boundaries of social formations and consequently accelerates the birth of pluralism. We witnessed the formation of cultural pluralism, linguistic pluralism, and racial pluralism throughout history, while human mobility plays a significant role in their composition. It is recognized that the communication and consensus processes in pluralistic societies are much more complicated than those in a unitary one. Thus, in this thesis, a computational approach is proposed to study such processes, and in particular, language evolution in a linguistic pluralism context is focused.

We consider a complex network where individuals are represented by nodes (agents) and the relations in between as edges. Then, we propose a novel naming game, called multi-language naming game (MLNG), to simulate the emergence of shared lexicons in a society via communication and consensus among agents. Unlike conventional naming games, which adopt a single-language context, the MLNG creates a multi-language context by defining agents as different-language speakers who cannot communicate with others without a translator (interpreter) in between. Simulation results reveal that the probabilities of different communication ways can be estimated based on the network features and the ratio of different language speakers. Also, the consensus convergence speed is found to have a power-law-like relationship with the proportion of translators.