Unsupervised spatial urban data representation learning
Urban environments are complex ecosystems comprising diverse spatial entities such as regions, road networks, and points-of-interest (POIs), which fundamentally shape our urban infrastructure and societal interactions. Understanding these spatial entities is crucial for various urban applications, f...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis-Doctor of Philosophy |
| Language: | English |
| Published: |
Nanyang Technological University
2024
|
| Subjects: | |
| Online Access: | https://hdl.handle.net/10356/179599 |
| _version_ | 1811692557750501376 |
|---|---|
| author | Zhang, Liang |
| author2 | Long Cheng |
| author_facet | Long Cheng Zhang, Liang |
| author_sort | Zhang, Liang |
| collection | NTU |
| description | Urban environments are complex ecosystems comprising diverse spatial entities such as regions, road networks, and points-of-interest (POIs), which fundamentally shape our urban infrastructure and societal interactions. Understanding these spatial entities is crucial for various urban applications, from optimizing transportation networks to informing urban planning and management decisions. In this regard, traditional methods relying on labor-intensive field surveys and questionnaires are gradually transformed by urban computing methods through integrating advanced machine learning techniques. However, these methods often focus on single urban application and require extensive labeled data as supervision, limiting their real-world applicability. Leveraging the wealth of unlabeled urban data available from sources like OpenStreetMap (OSM), this thesis proposes to develop pretraining methods specifically tailored for unlabeled urban data and investigate spatial entity representation learning problem within this context, aiming to address the scarcity of labeled urban data and facilitate the application of learned representations across diverse urban challenges. Specifically, in this thesis, we focus on three essential spatial entities in urban environments, namely region, road network, and points-of-interest (POI).
Effective representation learning of urban regions from the widespread multi-view urban data is crucial for various applications. However, existing methods often struggle with extracting representations within each single view and integrating them across multiple views effectively. In response to these challenges, we propose ReMVC (Region Embedding with Multi-View Contrastive Learning), a novel framework for region representation learning. ReMVC employs a hierarchical contrastive learning framework to extract representations within each view and fuse representations across views. Specifically, it emphasizes the comparison of different regions within each view to capture distinctive properties and promotes cross-view correlations by comparing the same region across different views. Experimental results demonstrate the effectiveness of ReMVC in two urban applications.
Road network serves as a foundational infrastructure supporting various real-life applications, yet raw road network data cannot be directly utilized by machine learning models designed for these tasks. Accordingly, road network representation learning (RNRL) is studied by learning vector representations of roads. Despite recent successes in adopting graph neural networks (GNNs) for RNRL, these models still face critical challenges in capturing high-order and long-range relationships between roads. To tackle these challenges, we introduce HyperRoad (Hyper}graph-Oriented Road network representation), a novel framework that leverages the power of hypergraphs for road network modeling. HyperRoad addresses the limitations of existing models by performing hypergraph-based dual-channel aggregation to capture high-order relationships among roads, as well as designing a novel hyperedge classification task to capture long-range relationships. Experimental results demonstrate that HyperRoad achieves impressive improvements compared to existing baselines across five applications.
The proliferation of location-based services has revolutionized user experiences, with recommendation systems playing a pivotal role in suggesting users Points-of-Interest (POIs). Despite their success, contemporary POI recommendation systems often necessitate extensive user-POI interaction data to learn high-quality POI representations and encounter critical cold start dilemmas. Consequently, learning to generate warm-up representations for these cold start POIs becomes imperative. Existing methods relying on crowdsourcing or adaptation from cold start product recommendation often suffer from low-fidelity sample issues and overlook rich spatial context information associated with POIs. To address these challenges, we propose a novel framework named DiffPOI (Diffusion-Oriented POI representation). DiffPOI treats POI representations as images and employs diffusion models to generate representations conditioned on rich POI semantic and spatial contexts. Experimental results show the superiority of our approach over existing methods, particularly in cold start POI recommendation applications.
In summary, this thesis presents several novel methodologies to address urban spatial entity representation learning problems. The proposed solutions have demonstrated improved performance across various scenarios and urban applications, offering valuable tools for analyzing large-scale unlabeled urban data. |
| first_indexed | 2024-10-01T06:37:41Z |
| format | Thesis-Doctor of Philosophy |
| id | ntu-10356/179599 |
| institution | Nanyang Technological University |
| language | English |
| last_indexed | 2024-10-01T06:37:41Z |
| publishDate | 2024 |
| publisher | Nanyang Technological University |
| record_format | dspace |
| spelling | ntu-10356/1795992024-09-04T07:56:36Z Unsupervised spatial urban data representation learning Zhang, Liang Long Cheng School of Computer Science and Engineering Data Management and Analytics Lab c.long@ntu.edu.sg Computer and Information Science Data mining Representation learning Urban environments are complex ecosystems comprising diverse spatial entities such as regions, road networks, and points-of-interest (POIs), which fundamentally shape our urban infrastructure and societal interactions. Understanding these spatial entities is crucial for various urban applications, from optimizing transportation networks to informing urban planning and management decisions. In this regard, traditional methods relying on labor-intensive field surveys and questionnaires are gradually transformed by urban computing methods through integrating advanced machine learning techniques. However, these methods often focus on single urban application and require extensive labeled data as supervision, limiting their real-world applicability. Leveraging the wealth of unlabeled urban data available from sources like OpenStreetMap (OSM), this thesis proposes to develop pretraining methods specifically tailored for unlabeled urban data and investigate spatial entity representation learning problem within this context, aiming to address the scarcity of labeled urban data and facilitate the application of learned representations across diverse urban challenges. Specifically, in this thesis, we focus on three essential spatial entities in urban environments, namely region, road network, and points-of-interest (POI). Effective representation learning of urban regions from the widespread multi-view urban data is crucial for various applications. However, existing methods often struggle with extracting representations within each single view and integrating them across multiple views effectively. In response to these challenges, we propose ReMVC (Region Embedding with Multi-View Contrastive Learning), a novel framework for region representation learning. ReMVC employs a hierarchical contrastive learning framework to extract representations within each view and fuse representations across views. Specifically, it emphasizes the comparison of different regions within each view to capture distinctive properties and promotes cross-view correlations by comparing the same region across different views. Experimental results demonstrate the effectiveness of ReMVC in two urban applications. Road network serves as a foundational infrastructure supporting various real-life applications, yet raw road network data cannot be directly utilized by machine learning models designed for these tasks. Accordingly, road network representation learning (RNRL) is studied by learning vector representations of roads. Despite recent successes in adopting graph neural networks (GNNs) for RNRL, these models still face critical challenges in capturing high-order and long-range relationships between roads. To tackle these challenges, we introduce HyperRoad (Hyper}graph-Oriented Road network representation), a novel framework that leverages the power of hypergraphs for road network modeling. HyperRoad addresses the limitations of existing models by performing hypergraph-based dual-channel aggregation to capture high-order relationships among roads, as well as designing a novel hyperedge classification task to capture long-range relationships. Experimental results demonstrate that HyperRoad achieves impressive improvements compared to existing baselines across five applications. The proliferation of location-based services has revolutionized user experiences, with recommendation systems playing a pivotal role in suggesting users Points-of-Interest (POIs). Despite their success, contemporary POI recommendation systems often necessitate extensive user-POI interaction data to learn high-quality POI representations and encounter critical cold start dilemmas. Consequently, learning to generate warm-up representations for these cold start POIs becomes imperative. Existing methods relying on crowdsourcing or adaptation from cold start product recommendation often suffer from low-fidelity sample issues and overlook rich spatial context information associated with POIs. To address these challenges, we propose a novel framework named DiffPOI (Diffusion-Oriented POI representation). DiffPOI treats POI representations as images and employs diffusion models to generate representations conditioned on rich POI semantic and spatial contexts. Experimental results show the superiority of our approach over existing methods, particularly in cold start POI recommendation applications. In summary, this thesis presents several novel methodologies to address urban spatial entity representation learning problems. The proposed solutions have demonstrated improved performance across various scenarios and urban applications, offering valuable tools for analyzing large-scale unlabeled urban data. Doctor of Philosophy 2024-08-13T01:13:40Z 2024-08-13T01:13:40Z 2024 Thesis-Doctor of Philosophy Zhang, L. (2024). Unsupervised spatial urban data representation learning. Doctoral thesis, Nanyang Technological University, Singapore. https://hdl.handle.net/10356/179599 https://hdl.handle.net/10356/179599 10.32657/10356/179599 en This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License (CC BY-NC 4.0). application/pdf Nanyang Technological University |
| spellingShingle | Computer and Information Science Data mining Representation learning Zhang, Liang Unsupervised spatial urban data representation learning |
| title | Unsupervised spatial urban data representation learning |
| title_full | Unsupervised spatial urban data representation learning |
| title_fullStr | Unsupervised spatial urban data representation learning |
| title_full_unstemmed | Unsupervised spatial urban data representation learning |
| title_short | Unsupervised spatial urban data representation learning |
| title_sort | unsupervised spatial urban data representation learning |
| topic | Computer and Information Science Data mining Representation learning |
| url | https://hdl.handle.net/10356/179599 |
| work_keys_str_mv | AT zhangliang unsupervisedspatialurbandatarepresentationlearning |