Urban mobility analytics: A deep spatial–temporal product neural network for traveler attributes inference
Introduction
With the accelerating urbanization, around of the world’s population is expected to live in cities by 2050. Urban public transport systems (buses, trains, and ferries, etc.) serve a large number of passengers on a daily basis and play an important role in metropolitan areas. However, current public transport system/service designs are often capacity-maximizing while individual preference is only considered to a limited extent. There is a growing trend to allow a more user-centered public transport system, which better accommodates, e.g., different age groups and the disabled. This requires improvements from at least the following aspects: infrastructures/equipment (vehicles, stations, access facilities, etc.); operation (line planning and scheduling); added services (connection information, entertainment TV programs, and advertisements). To provide such a user-centered public transport system, where individual preference is well accommodated, a critical input is the attributes of travelers. A conventional way is to conduct travel surveys to identify individual travel patterns and preferences, which can be costly and time-consuming. For example, Shiftan et al. (2008) proposed to identify travelers’ behaviors with the data collected from the Utah Transit Authority household survey. Differently, this study develops methods to infer traveler attributes (e.g., age groups, residential areas) based on individual travel trajectories while avoiding costly and time-consuming large-scale surveys. In particular, we test the effectiveness of the proposed model, a new hybrid spatial–temporal correlation model, for inferring age groups and residential areas of passengers. This has the potential to be utilized to improve transit services to accommodate requirements and preferences of different passenger groups (e.g., the elderly may be less demanding in terms of travel time, while they prefer quiet buses or trains; and young commuters/workers with work trips may be more demanding in terms of travel time and reliability). Moreover, the proposed method for inference may also be used to recover missing information/labels associated with individuals in a dataset. Besides, the generated insights from this study on how to link personal attributes/information with observable travel trajectories and mobility patterns may also be incorporated in other application domains with spatio-temporal complexity.
Location information with different levels of time resolution can be regarded as human trajectories with different sampling rates. The digital human trajectories collected by different types of data (e.g., GPS positions of mobile phones, smart card data, etc.) can be utilized to analyze human travel behaviors and mobility patterns (Song et al., 2010) and further infer the attributes of travelers. For example, Gonzalez et al. (2008) indicated that human trajectories exhibit a high degree of temporal-spatial regularity. Therefore, some statistical models and traditional machine learning methods were utilized to uncover travel patterns of passengers based on individual trajectories. Recently, the travel pattern regularity of public transport was analyzed by using the rough set theory and K-Means++ in Ma et al. (2013) without considering the spatial patterns of transit riders. Sun and Axhausen (2016) applied the probabilistic factorization framework to reveal spatial–temporal patterns of urban mobility, which repeats the estimation process to improve solution quality. Further studies indicated that the mobility patterns extracted from human trajectories are related to passenger attributes (Zhong et al., 2015, Luo et al., 2016, Olmos et al., 2018, Li et al., 2019). Although some researchers studied the spatial and/or temporal travel behaviours based on human trajectories, and analyzed the relations between these behaviours and traveler attributes by traditional clustering methods, e.g., Density-Based Spatial Clustering of Applications with Noise (DBSCAN) and K-means (Mohamed et al., 2016, Kieu et al., 2014), the complex spatio-temporal inter-correlation among features of mobility patterns have not yet been fully studied and uncovered to infer traveler attributes.
In this study, we propose to uncover mobility patterns of passengers by inferring passenger attributes in public transport systems with the help of large-scale smart card usage data and land use data. We focus on identifying the passengers into three age groups (i.e., adults, seniors, and children) and inferring the residential areas of passengers. In particular, we will take the age group as an example to present the critical features of mobility patterns including both the spatial and temporal information together with an analysis of their relationships. Based on the extracted mobility features and analysis, a hybrid neural network consisting of two components is developed for traveler attributes inference. The first component, the Product-based Spatial–Temporal Module (PSTM), is used to analyze and capture the spatial–temporal correlations from the extracted features, where we tested two specific modules, i.e., Inner-Product-based Module (Inner-PNN) and Outer-Product-based Module (Outer-PNN). The second component, Compression Module (CM), is utilized for compressing the transit stop sparse matrix (reflecting spatial patterns of trips), and further extracting useful information and learning the embedding vectors from this matrix, where we also tested two specific models, i.e., the fully connected layers (FCLs) and Auto-Encoder-based Compression Module (AECM).
The main contributions of this paper are summarized in the following. (i) We uncover and extract representative spatial and temporal passenger behavior patterns from a large-scale real-world dataset collected in the largest metropolitan area in Australia (Greater Sydney area). To provide empirical insights regarding mobility patterns associated with different attributes of travelers, we use age group information as an example and quantify the correlations/mapping between the mobility patterns and age groups. The travel pattern analysis is further enhanced by utilizing land use information (Point of Interest or PoI), which enriches the analysis to emphasize both the temporal and spatial dimensions. (ii) To the best of our knowledge, this paper is among the earliest to illustrate the potential of inferring individual attributes from observable trajectories based on smart card data of public transport with deep-learning-based methods. In this context, we propose a hybrid Neural Network, which combines PSTM and CM for the age group and residential area inference. The developed approach can also be utilized to either infer or recover unknown or missing attributes or labels in a dataset, as will be further discussed in Section 6.1. (iii) We evaluate the proposed method on a large-scale real-world dataset collected in the largest metropolitan area in Australia (Greater Sydney area) and demonstrate the effectiveness of the proposed method against several baselines and state-of-the-art methods.
The rest of this paper is organized as follows. First, we review some related works in Section 2. Then, we provide detailed introduction to the dataset used, the descriptive data statistics, mobility feature representation, and travel pattern analysis in Section 3. Section 4 presents the proposed inference model and related techniques. The test and evaluation of the proposed method and comparison to other methods are presented in Section 5. Section 6 discusses the potential applications and implications from this study and future research directions, and then concludes the paper.
Section snippets
Related work
In this section, we review works related to this study from two aspects: travel behavior and individual attributes studies with different data sources; and the inference and mining/learning strategies of travel patterns and/or traveler attributes.
(Travel behavior and individual attribute studies) A branch of studies on travel behaviors are mainly based on surveys. For instance, Axhausen et al. (2002) collected 6-week continuous data from a travel diary survey to show the dynamic changes of
Data description and behavioural features
This section describes the real-world dataset from Sydney used in this study and mobility feature extraction details. Then, spatial and temporal mobility features are analyzed based on age groups as an example to illustrate the mapping between mobility patterns and individual attributes.
Methodology
We now present the proposed modeling framework of the hybrid neural network incorporating spatial–temporal dependencies to infer the traveler attributes. The overall architecture of the proposed model is depicted in Fig. 9, where we have two parallel sub-networks: (i) a Product-based Spatial and Temporal Module (PSTM) with an inner-product operation (Inner-PNN) for capturing spatial–temporal dependencies from the extracted features introduced in Section 3.2; (ii) an Auto-Encoder-based
Experiments
In this section, we first introduce the experiment settings and evaluation metrics. Then, we present the experiment results from three perspectives: overall comparison, network architecture analysis, and ablation study. In particular, the proposed model is compared with ten existing strategies, i.e., LDA, QDA, SVM, Ada, DT, XGBoost, MLP, PPC, C2AE and DeepSD. Also, we test different models for the two sub-networks defined in Section 4 (network architecture analysis). Moreover, we conduct a
Discussion
This study demonstrates the potential of inferring or recovering traveler attributes or labels based on the observed trajectories of travelers in public transit systems. Moreover, this study provides new perspectives on utilizing public transport data to understanding travel behavior patterns, which may be further utilized to improve public mobility services and added services. Several potential applications or implications from this study are briefly discussed below.
- •
Inferring other attributes:
CRediT authorship contribution statement
Can Li: Conceptualization, Investigation, Methodology, Writing - original draft, Writing - review & editing. Lei Bai: Conceptualization, Methodology, Writing - review & editing. Wei Liu: Conceptualization, Funding acquisition, Investigation, Methodology, Project administration, Writing - review & editing. Lina Yao: Conceptualization, Writing - review & editing, Methodology. S. Travis Waller: Conceptualization, Writing - review & editing.
Acknowledgments
We would like to thank the handling editor, Prof. Zhen (Sean) Qian, and the anonymous referees for their constructive comments, which have helped improve both the technical quality and exposition of this paper substantially. Dr. Wei Liu thanks the funding support from the Australian Research Council through the Discovery Early Career Researcher Award (DE200101793).
References (59)
- et al.
City attachment and use of urban services: benefits for smart cities
Cities
(2016) - et al.
Statistical methods versus neural networks in transportation research: Differences, similarities and some insights
Transp. Res. Part C: Emerg. Technol.
(2011) - et al.
Short-term forecasting of passenger demand under on-demand ride services: a spatio-temporal deep learning approach
Transp. Res. Part C: Emerg. Technol.
(2017) - et al.
Behavioural data mining of transit smart card data: a data fusion approach
Transp. Res. Part C: Emerg. Technol.
(2014) - et al.
Forecasting short-term subway passenger flow under special events scenarios using multiscale radial basis function networks
Transp. Res. Part C: Emerg. Technol.
(2017) - et al.
The geography of human activity and land use: a big data approach
Cities
(2020) - et al.
Combining smart card data and household travel survey to analyze jobs–housing relationships in beijing
Comput. Environ. Urban Syst.
(2015) - et al.
Explore spatiotemporal and demographic characteristics of human mobility via twitter: a case study of chicago
Appl. Geogr.
(2016) - et al.
Estimating multi-year 24/7 origin-destination demand using high-granular multi-source traffic data
Transp. Res. Part C: Emerg. Technol.
(2018) - et al.
Understanding commuting patterns using transit smart card data
J. Transp. Geogr.
(2017)
Long short-term memory neural network for traffic speed prediction using remote microwave sensor data
Transp. Res. Part C: Emerg. Technol.
Mining smart card data for transit riders’ travel patterns
Transp. Res. Part C: Emerg. Technol.
Smart card data use in public transit: a literature review
Transp. Res. Part C: Emerg. Technol.
Transit market research using structural equation modeling and attitudinal market segmentation
Transp. Policy
Understanding urban mobility patterns with a probabilistic tensor factorization framework
Transp. Res. Part B: Methodol.
Bayesian committee of neural networks to predict travel times with confidence intervals
Transp. Res. Part C: Emerg. Technol.
Inferring demographics from human trajectories and geographical context
Comput. Environ. Urban Syst.
Human mobility and socioeconomic status: Analysis of singapore and boston
Comput. Environ. Urban Syst.
A deep learning approach to real-time parking occupancy prediction in transportation networks incorporating multiple spatio-temporal data sources
Transp. Res. Part C: Emerg. Technol.
Observing the rhythms of daily life: a six-week travel diary
Transportation
Spatio-temporal graph convolutional and recurrent networks for citywide passenger demand prediction
Stg2seq: spatial-temporal graph to sequence model for multi-step passenger demand forecasting
Passenger demand forecasting with multi-task convolutional recurrent neural networks
Estimating origin-destination flows using mobile phone location data
IEEE Pervasive Comput.
Content-aware hierarchical point-of-interest embedding model for successive poi recommendation
Smote: synthetic minority over-sampling technique
J. Artif. Intell. Res.
Xgboost: A scalable tree boosting system
Travel demand prediction using deep multi-scale convolutional lstm network
Approaches to privacy preservation in intelligent transportation systems and vehicle–infrastructure integration initiative
Transp. Res. Rec.
Cited by (20)
Understanding the timing of urban morning commuting trips on mass transit railway systems
2024, Transportation Research Part C: Emerging TechnologiesIncome estimation based on human mobility patterns and machine learning models
2024, Applied GeographyResearch trends, themes, and insights on artificial neural networks for smart cities towards SDG-11
2023, Journal of Cleaner ProductionCharacterizing the activity patterns of outdoor jogging using massive multi-aspect trajectory data
2022, Computers, Environment and Urban SystemsCitation Excerpt :In contrast, our work focuses on capturing the interdependency between different aspects and detecting multiplex combined patterns from multi-aspect trajectory. By segmenting joggers into groups with similar activity patterns, much latent valuable information such as habits or preferences (Bajaj & Singh, 2019; Shen & Cheng, 2016), life styles (Ou & Cai, 2018) and socio-demographics can be revealed (Li, Bai, Liu, et al., 2021; Zhang, Aslam, Lai, et al., 2020). Recently, more studies about user segmentation and pattern characterizing have been made based on different data (Bajaj & Singh, 2019; Crawford, 2020; Egu & Bonnel, 2020; Ou & Cai, 2018).
STGNN-TTE: Travel time estimation via spatial–temporal graph neural network
2022, Future Generation Computer SystemsCitation Excerpt :Based on these models, more and more improved variants were presented. For instance, many works involved attention mechanism [34–38], adversarial loss [39–41] and some external information [36,42–46] to enhance the capability of spatial–temporal representation learning. In recent years, with the development of graph convolutional networks (GCN), deep learning methods based on spatial–temporal graphs began to be involved into traffic prediction.
A multi-task memory network with knowledge adaptation for multimodal demand forecasting
2021, Transportation Research Part C: Emerging TechnologiesCitation Excerpt :As discussed in Section 1, providing high-quality demand forecasting of station-sparse and data-scarce sources only based on their own information is challenging and the performance/accuracy is bounded by the level of information that is available. Therefore, some studies utilized alternative data sources to substitute the low-coverage/quality data as the expansion of massive transport data accessible in the digital era (Li et al., 2021). For example, since the insufficient sensor coverage across the traffic network limited the obtaining of traffic flow information comprehensively, Zhang et al. (2020b) adopted crowd-sourcing floating car data with a high coverage rate for flow estimation.