Dissertation
Estimating human mobility responses to social disruptions through spatio-temporal deep generative learning methods
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2024
DOI: 10.25820/etd.007615
Abstract
Estimating human mobility is an important task in diverse societal domains, including public health, public safety, transportation, agriculture, environmental science, etc. This thesis seeks to formulate concepts and develop methods that facilitate the estimation of human mobility responses to social disruptions. Social disruption refers to a significant alteration or upheaval in the patterns of social behavior and interactions within a society, such as disease outbreaks, natural disasters, wars, or other significant events. In the case of COVID-19, social disruption can be seen in the significant changes that have occurred in the way that people live and interact with one another. For example, the COVID-19 pandemic has caused significant social disruption by changing human mobility patterns, social interactions, and economic activity. Many countries have implemented social distancing measures, such as lockdowns, restrictions on gatherings, and closure of non-essential businesses, to slow the spread of the virus. Estimating human mobility to the COVID-19 pandemic is crucial since these measures have led to significant disruptions in people's daily lives, including changes in work arrangements, school closures, and limitations on travel. This thesis proposal aims to study the estimation or prediction techniques required for human mobility responses to such social disruptions using spatio-temporal data, the thesis takes COVID as an example, but the proposed work can be generalized to other types of social disruptions.
Human mobility response is the pattern of movement and migration among individuals and populations that adjusts their movements in response to changing circumstances, such as environmental factors or societal changes. It is crucial for policymakers and urban planners to evaluate different possibilities of human mobility responses during decision-making and resource allocation. Human mobility response estimation or prediction techniques required by these domains are challenging due to the spatio-temporal data non-stationarity, complex social contexts, and limited training data. As a result, spatio-temporal properties are explicitly needed to control or model in estimation model frameworks. To address the challenges of spatio-temporal heterogeneity and dependencies, extensive research has been conducted by data mining and geo-spatial communities. This thesis tackles these challenges specifically within the realm of human mobility response estimation.
Traditional methods for estimating human mobility responses often rely on statistical models that struggle to address real-world challenges in mobility estimation and simulation. These challenges include the effects of unknown, uncertain, and random factors. Moreover, these models typically analyze mobility changes on a city or country scale, which can be limiting. Recently, deep learning techniques have been developed for spatio-temporal data mining. For instance, Long Short-Term Memory networks (LSTMs) have been widely used in traffic accident prediction and flow prediction due to their ability to capture spatio-temporal correlations, resulting in improved prediction accuracy. However, most techniques for spatio-temporal task estimation use stationary predictors, which means they produce the same results from two runs on the same data. These methods do not account for unknown factors in predictions and often rely on large datasets. In the literature, this problem is often approached as a traffic demand prediction problem, which typically does not consider spatio-temporal correlations simultaneously. This thesis aims to leverage deep spatio-temporal models to estimate human mobility responses, with a particular focus on addressing unknown factors in real-world estimation tasks.
While human mobility responses represent the natural distribution of human activities, estimating these natural mobility patterns for specific regions presents several unique challenges. First, human mobility responses depend on numerous complex social and physical factors, some of which are known, while others remain unknown and have limited data available. Second, human mobility responses often exhibit spatio-temporal non-stationarity, characterized by complex non-linear interactions between time and space. Third, a distinctive aspect of spatio-temporal problems, setting them apart from classical data-mining domains, is the presence of dependencies among measurements induced by the spatial and temporal dimensions.
This thesis presents three major contributions. First, the proposed COVID-GAN is a conditional generative adversarial network that estimates human mobility response under a set of social and policy conditions integrated from multiple data sources in one city. Second, to estimate the human mobility spatio-temporal dependencies across cities and time periods, this thesis presents STORM-GAN, which uses a meta-learning-based generative model and is facilitated by a novel spatio-temporal task-based graph (STTG) embedding. Finally, this thesis presents LAS-GAN, a spatio-temporal generative model that integrates language-augmented semantic context to represent spatial variability to enhance human mobility estimation performance. Comprehensive evaluations using real-world datasets demonstrate meaningful results and superior performance compared to baseline methods.
Details
- Title: Subtitle
- Estimating human mobility responses to social disruptions through spatio-temporal deep generative learning methods
- Creators
- Han Bao
- Contributors
- Juan Pablo Hourcade (Advisor)Bijaya Adhikari (Committee Member)Caglar Koylu (Committee Member)Yanhua Li (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Informatics
- Date degree season
- Summer 2024
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007615
- Number of pages
- xiv, 158 pages
- Copyright
- Copyright 2024 Han Bao
- Language
- English
- Date submitted
- 07/15/2024
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 136-176).
- Public Abstract (ETD)
- Human mobility refers to the movement and migration patterns of people. Under- standing these patterns offers valuable information for public health, safety, and urban plan- ning. Estimating human mobility becomes crucial during significant social disruptions like the COVID-19 pandemic, which drastically altered movement patterns due to lockdowns and social distancing measures. Traditional methods struggle with accuracy due to un- known factors and the complexity of social contexts. This thesis leverages advanced deep learning techniques to address these challenges, introducing three innovative models. The COVID-GAN model estimates human mobility under different social and policy conditions using data from one city. The STORM-GAN model captures spatio-temporal dependencies across different cities and times using meta-learning and graph embeddings. The LAS-GAN model incorporates language-based semantic context to improve spatial variability representation and enhance mobility estimation. These models aim to provide better estimates of human mobility in response to social disruptions, aiding policymakers and planners in making informed decisions. Evaluations using real-world data demonstrate superior performance compared to traditional approaches.
- Academic Unit
- IDGP in Informatics
- Record Identifier
- 9984698053802771
Metrics
12 File views/ downloads
4 Record Views