HUMAN MOBILITY ANALYTICS WITH DEEP REPRESENTATIONS

Join Zoom Meeting
https://us02web.zoom.us/j/88680836039

Meeting ID: 886 8083 6039

Abstract:

One of the largest dynamics of our society is the human mobility whose pricelessness is concealed by its mighty complexity and diversity, in both spatial and temporal aspects. A better understanding of such dynamics can benefit our livings in various perspectives. This is the reason why human mobility analytics has been an important subject intriguing a large bulk of research spanning across different areas. Various technologies are developed to understand, model and exploit human moving patterns, among which one common prerequisite is an appropriate way to represent human mobility such that the representation is expressive and flexible enough to capture the richness of geographic and semantic attributes. Despite enormous efforts invested, traditional studies in this pursuit remain insufficient to account for the complexity and diversity of human mobility. Thanks to the increasing availability of human trajectory data and computation resources, people can extract more abstracted and distributed representations for human mobility by employing more sophisticated transformation and deeper modeling.

In this thesis, using deep representation techniques, we conduct systematic analytics on human mobility by stratifying our view into three representation levels: location, trajectory, and aggregation level, each with corresponding analytic tasks to be addressed.

At the location level, we target learning location embeddings with better semantic coherence in an unsupervised learning paradigm. Through disentangling the semantic subspace from the geographic subspace, we present a simple yet effective technique to improve the generalizability of the learned embeddings, whose effectiveness is verified through extensive experiments and case studies.

At the trajectory level, we attack the problem of human trajectory synthesis. In this work, a generative adversarial training scheme is leveraged to learn a non-parametric generator model. This overcomes the difficulty in finding a tractable approximation for the joint probability distribution of traces and preserves faithful geographic and semantic attributes of original data.

At the aggregation level, we are tasked to infer fine-grained citywide human flow given coarse-grained input so as to save cost for maintaining modern urban flow monitoring systems. We first present a model entitled UrbanFM which resolves the challenges originate from structural constraints and external complexities of the problem. Constructing upon the first version, we further expose UrbanPy, an enhanced model characterized by its pyramid structure which performs large-scale upsampling more effectively.