Files and links (1)
BANG AN Thesis Revised15.58 MB
Embargoed Access, Embargo ends: 08/30/2026
Dissertation
Machine learning solutions to urban incident forecasting and understanding on spatiotemporal data
University of Iowa
Doctor of Philosophy (PhD), University of Iowa
Summer 2024
DOI: 10.25820/etd.007767
Abstract
The aim of this thesis is to investigate novel machine learning and artificial intelligence (AI) solutions to the problem of urban incident forecasting and understanding based on big spatiotemporal urban data. Urban incidents are defined as abnormal events such as traffic accidents and crimes occurring over different time and locations. It is important for city management to identify and understand such incidents, including when, where, and why these incidents may happen. The ability to detect urban incidents at an early stage facilitates traffic congestion, enhances public safety, optimizes resource allocation, and reduces property losses. More importantly, understanding incident patterns and explaining the relationships between the occurrence of incidents and associated socio-environmental brings a possibility to prevent incidents from happening. With the availability of urban big data such as people's GPS trajectory and weather conditions advances in deep learning techniques, AI models are promising solutions to the problem and have the potential to transform the practices in urban administration and safety .
Forecasting urban incidents is challenging due to the spatial heterogeneity of incident patterns, which vary dramatically over different regions. A performing model learned in a region might become less-satisfying in another area. Furthermore, being able to prioritize the most important incident is also non-trivial. Because widely used optimization objectives such as cross-entropy are designed to make accurate predictions for all incidents, precision on the most important ones is less considered. Lastly, understanding the urban incident patterns and relationships to socio-environmental factors is challenging due to complex non-linear dependencies and the non-intuitive nature of spatiotemporal data. Understanding the reasoning of the models can help explain the forecasting results.
Classic machine-learning methods such as regression-based and tree-based methods have difficulty in learning from abundant spatial-temporal data and capturing complex incident patterns. Recently, some researchers have introduced deep learning techniques to improve prediction accuracy. Nevertheless, the performance is limited due to the unique pattern of traffic accidents including unbalanced data and heterogeneous spatial distribution. Prior works addressing spatial heterogeneity primarily rely on using ensembles of multiple models capturing various patterns on different sub-regions of the study area. However, the sub-regions are predefined by the user's domain knowledge and it may not lead to optimal performance. Moreover, no existing work tries to prioritize the most important incidents by directly optimizing the model for ranking quality. This thesis proposes the first solution to forecast the rank of urban incidents and optimize their ranking quality. Furthermore, existing methods interpret urban incident patterns by exploring the relationships between the occurrence of incidents and socio-environmental attributes via linear models. No existing work tries to develop an intrinsically interpretable spatiotemporal model.
This thesis presents the HinNet and LISA to systematically address the spatial heterogeneity issues. HintNet employs a spatial partitioning algorithm to group regions with similar incident patterns, and then learns multiple models over partitions. The LISA framework integrates spatial partitioning and model training in a way that heterogeneous patterns are captured automatically. To rank the urban incidents, this thesis proposes SpatialRank using a novel ranking objective function to optimize predictive ranking quality. Comprehensive experiments on both car accident and crime datasets demonstrate superior performance over baselines by up to 12.7%. To interpret deep learning models learned on spatiotemporal datasets, this thesis proposes GeoPro-Net, an intrinsically interpretable spatiotemporal model for spatiotemporal forecasting problems. GeoPro-Net learns different sets of prototypes of concepts inherently, and projects them to real-world cases for interpretation. Comprehensive evaluations using real-world datasets and case studies demonstrate better interpretability while still achieving competitive prediction performance compared with other baselines.
Details
- Title: Subtitle
- Machine learning solutions to urban incident forecasting and understanding on spatiotemporal data
- Creators
- Bang An
- Contributors
- Weiguo Fan (Advisor)Xun Zhou (Committee Member)Kang Zhao (Committee Member)Nick Street (Committee Member)
- Resource Type
- Dissertation
- Degree Awarded
- Doctor of Philosophy (PhD), University of Iowa
- Degree in
- Business Administration
- Date degree season
- Summer 2024
- Publisher
- University of Iowa
- DOI
- 10.25820/etd.007767
- Number of pages
- xii, 112 pages
- Copyright
- Copyright 2024 Bang An
- Language
- English
- Date submitted
- 07/02/2024
- Description illustrations
- illustrations, tables, graphs
- Description bibliographic
- Includes bibliographical references (pages 99-112).
- Public Abstract (ETD)
- Urban incidents, such as traffic accidents and crimes, pose significant challenges for city management. This thesis explores advanced machine learning and artificial intelligence (AI) solutions to enhance the forecasting and understanding of urban incidents using big spatiotemporal urban data. Early detection and understanding of these incidents can offer practical tools for city management to enhance urban safety and efficiency. However, it is computationally difficult to capture rare urban incidents over regions with different incident patterns. Additionally, it is challenging to understand the reasons leading to incidents. Existing machine learning solutions fail to capture complicated incident patterns over time and space and are struggling with explaining their predictions. This thesis presents our novel machine learning solutions to answering when, where, and why these incidents may happen. First, we propose HintNet and LISA framework to capture diverse incident patterns over space. Second, we propose SpatialRank to identify the riskiest incidents in a study area. Lastly, we propose GeoPro-Net to explain why there may be an incident. Comprehensive experiments on multiple large-scale crime and traffic accident datasets show that our solutions can accurately forecast urban incidents and explain why they happen.
- Academic Unit
- Bus Admin Graduate Programs
- Record Identifier
- 9984698054202771
Metrics
5 Record Views