Cycling provides considerable environmental and societal benefits. However, bicyclists are among the most vulnerable road user groups. To develop effective mitigation measures that can reduce bicycle injuries, it is necessary to identify the factors that affect the injury severity of bicycle crashes. These factors are often aggregated at different scales, and their associations with bicycle injury risk are complex. It is essential to explore the spatio-temporal patterns of bicycle injuries and the potential factors. In conventional bicycle safety studies, influences of macro-level and micro-level factors on bicycle injury severity are often evaluated separately. Interdependence between macro-level and micro-level factors in the association between bicycle injury risk and potential factors is seldom explored. In this study, a dual-layer complex network model is proposed to integrate both macro- and micro-level factor sub-networks, model the changes in network topology over time, and capture the non-linear relationships between bicycle injury severity and potential factors. Bicycle crash data from 289 Lower Super Output Areas (LSOAs) in London from 2017 to 2019 are used to calibrate the model. Results indicate strong explanatory power of the proposed model under heterogeneous network conditions. At the micro-level, lighting condition significantly affects both slight injury and fatal and severe injury (FSI) risk, and speed limit remarkably interacts with several macro-level factors. At the macro-level, green area, non-residential building area, road area, and road density are significant contributory factors for both slight injury and FSI risk. In addition, traffic flow, average bus stop density, and household income significantly interact with several micro-level factors. Furthermore, age exhibits a strong connection with several macro-level and micro-level factors for FSI risk. Lastly, male proportion, population density, and green area exhibit strong linkage with the factor networks for slight injury risk. These findings should inform the implementation of future urban design and transport policy strategies that can reduce bicycle injury risk and promote cycling as a sustainable transportation mode in the long run.
Mobile phone signalling and other passive-sensing mobility data offer population coverage at low cost, but they record where people are rather than what they do. This semantic gap limits their value for activity-based travel behaviour analysis and planning diagnosis. To address this limitation, we propose UrbanAct-GPT, an evidence-constrained and task-adapted LLM framework that converts location-only signalling records and POI information into period-level activity types and day-level behavioural roles through Context-constrained Chain-of-Thought prompting, closed activity and role taxonomies, and QLoRA fine-tuning. On an independent test set, UrbanAct-GPT achieves 0.95 activity accuracy and 0.70 role accuracy. Benchmarking against both general-purpose LLMs and conventional activity-inference baselines shows that UrbanAct-GPT’s advantage lies mainly in category-balanced inference for context-dependent activities such as eating, shopping, leisure, care, and fitness, rather than routine home and work recognition alone. Applied to approximately 2.1 million anonymised users in Shanghai, UrbanAct-GPT transforms passive mobility traces into activity-specific and time-sensitive evidence. In Shanghai, the inferred activity semantics show that mobility concentrations can reflect different urban demands, including peak-period commuting around employment centres, late-return trips along cross-district corridors, and evening non-work activities in residential communities. In this sense, the model links mobility intensity to the urban functions supported by different places at different times of day. By distinguishing activity purposes across locations and time periods, the results support targeted transport service adjustment, community-facility provision, night-time amenity planning, and jobs–housing diagnosis. Overall, UrbanAct-GPT provides a pathway for recovering activity semantics from passive-sensing mobility data and supporting city-scale planning diagnosis.
Baidu Apollo Go, launched in Wuhan, marks the first large-scale commercial deployment of autonomous ride-hailing in China. Before such services are rolled out more broadly, there is an urgent need for an objective assessment of public acceptance. However, existing studies largely rely on hypothetical scenarios and short observation windows, so their findings may deviate from real-world dynamics. This study focuses on a full year following the surge in public attention surrounding Apollo Go and uses related comments posted on Weibo as the empirical basis to investigate the evolution of public sentiment orientation. Specifically, we collect 41,009 comments containing references to Apollo Go and autonomous ride-hailing from 7 July 2024 to 6 July 2025 via a web spider, apply Latent Dirichlet Allocation for topic modelling, and employ the Baidu Smart Cloud Sentiment Analysis API for sentiment classification. By situating these comments in the context of major events and regional differences, we explore how public attention and sentiment toward key topics evolve over time. The main findings are as follows. (1) Public attention exhibits a clear stage-wise shift. In the early stage, discussion mainly focuses on the employment impact on ride-hailing drivers, whereas, as commercial operations progress, public attention gradually shifts toward passenger safety and ride experience. (2) Although the overall sentiment score for the full study period is only 0.35, public sentiment exhibits a fluctuating upward trend over time. (3) Major events lead to pronounced fluctuations in public sentiment. Technological breakthroughs, disclosures of safety performance, and positive industry-development signals generally improve public sentiment, whereas safety accidents and corporate operational risks tend to reduce it. (4) Public sentiment also exhibits clear regional differences. Among provinces where autonomous ride-hailing services operate over an area of at least 100 km2, those with per capita GDP above USD 20,000 generally exhibit lower overall sentiment scores.
Public transit networks structure social and economic activities by bringing diverse populations together in shared spaces. This study examines the spatial relationship between faregate replacements and nearby crime patterns within the Washington Metropolitan Area Transit Authority (WMATA) Metrorail network in the District of Columbia, Montgomery County, and Prince George’s County, Maryland. Faregate replacements were implemented system-wide across all 98 Metrorail stations between July 2023 and September 2024; however, the rollout was staggered, and precise completion dates were not uniformly documented for all stations. This study therefore uses the seven stations with verified replacement dates as treatment stations and uses propensity score matching (PSM) to select seven control stations based on pre-intervention property and violent crime levels to assess changes in nearby crime patterns following the intervention. These stations constitute the earliest confirmed interventions, allowing for a before-and-after comparison and a quasi-experimental design with temporal precision. We use a spatial analytical approach employing the weighted displacement quotient (WDQ) and negative binomial difference-in-differences (DID) models to analyze property and violent crime patterns between the pre- and post-intervention periods (2022 and 2024). This approach examines Census blocks in street-network buffers of 0–400 m, 400–800 m, and 800–1,200 m around stations. The results indicate statistically significant reductions in property crime within the 0–400 m treatment area, with weaker effects at greater distances from stations. In contrast, violent crime exhibited no statistically significant changes following faregate replacement, suggesting that the effects of access-control interventions may differ across crime types. This study highlights the significance of considering interventions in the built environment and, more broadly, underscores the need for urban policy to account for the broader social consequences of transportation design.
Drone delivery, as a low-altitude logistics application of urban air mobility (UAM), may reshape last-mile distribution and urban service accessibility. However, its societal diffusion cannot be fully understood through conventional binary acceptance models, because consumers may differ substantially in their readiness to progress from awareness to repeated use. This study investigates heterogeneous consumer readiness for drone delivery using 1,018 valid survey responses collected in Shenzhen, China, where urban drone delivery has entered early deployment. Drawing on the Transtheoretical Model, adoption is conceptualized as an ordered five-stage readiness process: precontemplation, contemplation, preparation, action, and maintenance. An integrated framework combining the Technology Acceptance Model and the Theory of Planned Behavior, extended by environmental concern and perceived risk, is used to identify attitudinal latent constructs. A hybrid model integrating a Multiple Indicators and Multiple Causes model and a Partial Proportional Odds Model examines how socio-demographic attributes and psychological perceptions shape stage-specific readiness, with marginal effects quantifying heterogeneous influences across stages. Results show that drone delivery adoption is not a linear extension of general acceptance but a staged readiness process shaped by different psychological mechanisms. Socio-demographic attributes mainly explain baseline differences in adoption propensity, while attitudinal constructs exert stronger and more stage-dependent effects. Early-stage readiness is constrained by perceived risk and uncertainty, but is also sensitive to social influence and environmental concern; progression toward action and maintenance is more closely associated with perceived usefulness, ease of use, perceived behavioral control, attitude, and behavioral intention. This pattern indicates a pathway from risk reduction and cognitive activation to technology evaluation, experiential confidence, and behavioral consolidation. Findings provide behavioral evidence for stage-targeted risk communication, service design, and demand cultivation strategies in early urban drone delivery deployment.
This study links mobility styles with carsharing intentions and next-generation vehicle preferences in urban Japan. An online survey of licensed drivers in Tokyo Metropolis, Kanagawa Prefecture, and Osaka Prefecture captured mode use frequencies and stated preferences by travel purpose. Cluster analysis identified three mobility styles: private-car users, public-transport users, and a diversified, high-carsharing group. The diversified profile combined high carsharing participation with high private-car ownership, indicating that carsharing can complement rather than replace existing car access. Random Forest results indicate that age, car usage frequency, and income best distinguish carsharing users from non-users. Carsharing preferences varied substantially by stated-use context. For daily and business trips, most respondents prefer carsharing stations within 400 m and short distances; for tourism and hometown visiting, they accept longer access to carsharing stations and longer trips. Compact and Kei cars (Japanese light vehicles) in Japan dominate daily use, while minivans and SUVs gain importance for leisure. Gasoline and hybrid vehicles are most preferred, with limited willingness to pay extra for electric vehicles. The findings support segment- and purpose-specific carsharing station placement, fleet composition, and gradual electrification strategies in mature, transit-rich metropolitan areas.
Rising temperatures driven by climate change and development patterns affect travel behavior, including public transit use and active transport; however, most studies rely on ambient temperature measures and assume uniform behavioral responses. In this study, we investigated how individuals’ perceived heat thresholds are associated with transit use and moderate-to-vigorous physical activity (MVPA), among adults in a humid subtropical climate. We used data from a 2022 cross-sectional survey of adults in Austin, Texas (n = 1,087). Respondents self-reported their perceived heat threshold (i.e., the temperature at which they would not go on a walk outside), weekly frequency of using public transit, and MVPA levels via the International Physical Activity Questionnaire–Short Form. We used ordinal logistic regression to assess the association between perceived heat threshold and transit use (0, 1–2, ≥3 days), multinomial logistic regression to examine the association within transit use categories (0, 1–2, 3–4, ≥5 days), and multivariable linear regression to assess the association with MVPA. Individuals who reported higher perceived heat thresholds had higher odds of using public transit. This association was primarily driven by the transition from non-use to occasional transit use (1–2 days/week), with no evidence that perceived heat thresholds were associated with more frequent transit use. In contrast, perceived heat threshold was not significantly associated with MVPA. Our findings suggest that individual heat-related behavioral thresholds may help explain individual mobility decisions, but not overall physical activity levels.
The impact of weather on cycling demand has usually been investigated for day-to-day variations, often neglecting the impacts on individual-level habitual cycling behavioral changes over time. This study addresses this gap by developing a framework based on Bayesian online changepoint detection (BOCD). We first use the BOCD method to detect changes in habitual weekly cycling frequencies for each individual. Multinomial logit (MNL) models are then used to examine the effects of weather conditions and sociodemographics on the occurrence of habitual behavioral changes. To model the nonlinear effects of weather attributes, we propose an automatic procedure for ’optimal’ piecewise linear specifications instead of using predefined categories. We tested our approach on GPS datasets from Switzerland comprising 698 cyclists with observation periods from 31 to 79 weeks, combined with public weather data. Results show relatively stable cycling frequencies in our sample, with an average of 1.95 changepoints detected per individual. MNL models indicate that snowfall and precipitation significantly reduce habitual cycling frequency. Temperature shows a nonlinear relationship, with moderate temperatures triggering an increase in habitual cycling frequency, and more extreme conditions leading to a decline. Moreover, heterogeneity in weather response is observed, particularly with respect to e-bike ownership and cycling experience. This framework allows for consistent comparison of weather effects across regions and provides a basis for examining behavioral adaptation to a changing climate.
Health equity depends not only on insurance eligibility, but also on access to the administrative system that enables coverage. Using Medicaid administrative offices in the United States as a case, we assess accessibility from over 80,000 census tracts to the nearest Medicaid office across driving, public transit, bicycling, and walking using distances and travel times. We identify a proximity–access paradox: while geographic distances to offices are broadly similar across modes, functional accessibility diverges sharply once translated into time and transportation feasibility. Driving times are generally short and consistently feasible, whereas walking and transit often require substantially longer times, and public transit is infeasible for over 40% of tracts. Multilevel analyses further show that accessibility varies systematically with neighborhood income, racial composition, age structure, disability prevalence, and state context. These findings highlight transportation barriers as an important structural determinant of administrative access to health insurance coverage and suggest that improving health equity requires policy attention not only to eligibility and coverage expansion but also to the transportation systems that enable people to reach the administrative institutions of health systems.
Planned special events (PSEs) bring considerable economic benefits to cities but simultaneously challenge the performance of urban transportation systems. Existing studies focusing on PSEs either rely on survey data with limited population coverage or adopt aggregate analyses that cannot precisely identify PSE-related trips. Utilizing ride-hailing data with detailed textual addresses, this study accurately identifies concert-related trips in Tianjin, China, and investigates the spatiotemporal dynamics and determinants of ride-hailing demand using machine learning models and SHAP-based interpretation. A total of 7,022 ingress trips and 3,381 egress trips were identified. Approximately 70 % of ingress trips occur between one hour before venue opening and one hour before the concert starts, while around 70 % of egress trips are generated within one hour after the concert ends. In terms of travel distance, 86.1 % of ingress trips and 77.5 % of egress trips are concentrated within 12 km of the venue. Spatially, ingress trip origins cluster in the urban core and several prominent urban attractions, whereas egress trip destinations are oriented toward hotels, university districts, and train stations. Machine learning results indicate that concert-related ride-hailing demand is primarily shaped by location-related factors and built-environment characteristics, with clear phase-specific differences. Distance to Tianjin Olympic Center (TJOC) and Hotel Points of Interest (POIs) density consistently emerge as the most influential variables. The distance effect is stronger during ingress and weaker during egress, while Hotel POIs density maintains a positive association, exerting greater influence during egress. Population density exhibits a consistently negative relationship with ride-hailing demand. Parking POIs density, Restaurant POIs density, Bus stops density, and Land use diversity, display stage-dependent effects, jointly reflecting the influence of transport-related infrastructure and built environment on ride-hailing demand during large-scale events. These findings enhance understanding of PSE-related ride-hailing patterns and provide practical insights for proactive traffic management during large-scale events.
While transport research has shown that individual attitudes and subjective norms affect travel behavior, most studies overlook how household-level travel socialization processes (i.e., how attitudes and behaviors are transmitted between members) affect these patterns over time. The current study examines socialization processes in two household types, two-partner and family households, focusing on how fathers (or male partners), mothers (or female partners), and children influence each other’s travel attitudes and behaviors over time. Using a two-wave cross-lagged panel model from the Netherlands Mobility Panel, we test the directionality of these effects. Results indicate that (i) in two-partner households, sustainable modes show more bidirectional effects between partners, though train use remains asymmetrical; (ii) in family households, the socialization process is largely parent-driven, with mothers playing a central role in influencing sustainable travel attitudes/behaviors of both children and partners; (iii) car use shows bidirectional effects between children and parents; and (iv) in both household types, individual travel behaviors are also highly stable over time, indicating the importance of behavioral consistency in addition to socialization processes. Mode-specific strategies targeting influential household members; mothers for cycling, parents for train use, and children for car reduction; can more effectively promote sustainable travel.
Activity-based models (ABMs) traditionally rely on the “skeleton activity” assumption, typically ‘out-of-home work/school’ to organize daily activity participation. However, this assumption has largely been embedded procedurally and lacks empirical validation, particularly under the emergence of telecommuting and hybrid work. This study uses a 24-hour activity diary data from the 2023 British Columbia Activity and Time Use Survey (ATUS) to investigate whether skeleton activities emerge empirically and whether their identity varies across work arrangements. Separate Bayesian Network (BN) models are estimated for commuters, telecommuters, hybrid workers, non-workers, and students. This study compares the skeleton-free structures (PC, Greedy Thick Thinning) against skeleton-assumed structures (Augmented Naïve Bayes, Tree-Augmented Naïve Bayes) to evaluate whether a dominant organizing activity is supported by the data. Across all worker profiles, the Tree-Augmented Naïve Bayes (TAN) algorithm provides the best statistical fit, confirming the presence of a skeleton-type organization while allowing additional interdependencies among activities. However, the identity of the skeleton activity differs systematically by group: ‘Out-of-home work or school’ anchors commuters, hybrid workers, and students; ‘online work or school’ anchors telecommuters; and ‘out-of-home shopping’ anchors non-workers. Furthermore, the learned networks reveal distinct interdependencies across activity types and spatial contexts, reflecting differences in flexibility and constraints across populations. Socio-demographic factors, particularly household composition, vehicle availability, and license ownership, significantly influence activity participation and interact with skeletal structure. For example, individuals with children in the household are likelier to participate in ‘pick-up/drop-off activities. Out-of-sample validation confirms the stability of the inferred networks for most groups. The findings empirically challenge traditional commuter-centric skeleton assumptions by showing evidence that skeletal anchors are population specific and context-dependent, providing a foundation for more flexible and inclusive activity-based modeling systems.
Accurate Origin-Destination (OD) data are essential for understanding complex travel behavior and advancing sustainable, equitable urban bus systems, yet remain globally scarce. This study develops and validates a context-aware framework that reconstructs bus OD flows by integrating large-scale mobile trajectories with online Map API contextual information. The framework comprises three stages: trajectory extraction, bus travel mode identification, and stop- and route-level OD flow aggregation. Evaluation in Guangzhou demonstrated the framework’s efficacy and the critical role of Map API integration. An ablation study at both micro and macro levels quantified this contribution. The API-Model achieved an F1 Score of 82.42% for bus trip identification, substantially outperforming a topology-based baseline model. It also yielded strong correlation with official ridership data (R2 = 0.786). The reconstructed OD flows revealed demand complexities invisible to boarding-only data, including stop-level functional imbalances between trip generators and attractors, and Pareto-distributed ridership concentration. An illustrative network optimization application further quantified planning value. Compared with a boarding-based strategy, the OD-based approach improved OD demand coverage by 5.15 percentage points and alighting demand coverage by 16.61 percentage points. These findings support a paradigm shift towards demand-based transit planning grounded in observed travel behavior data.
Rapid urbanization and unprecedented population mobility in China have intensified the need to understand the drivers of migrants’ settlement decisions. Despite the central role of daily mobility in shaping urban experiences, little is known about how transportation convenience perceptions and travel modes influence long-term settlement intentions. This study addresses this gap by developing a multi-dimensional behavioral‑perceptual framework of transportation convenience, moving beyond single‑dimension perceptual or amenity‑based approaches. Crucially, it further examines how the effects of these perceptions vary across travel modes (walking, biking, metro, bus, and car), an aspect largely overlooked in prior research. Using the 2022 City Health Examination Survey, covering 240,177 migrants across 57 Chinese cities, and employing multi-level mixed-effects models, we examine the associations between transportation perceptions, travel modes, and settlement intentions. Findings show that more positive perceptions of transportation convenience are positively associated with settlement intentions, including the walk environment (β = 0.124, P < 0.01), riding environment (β = 0.090, P < 0.01), public transport interchanges (β = 0.131, P < 0.01), road congestion (β = 0.097, P < 0.01), and commuting time (β = 0.089, P < 0.01). The influence of transportation perceptions varies across travel modes, with active travellers (walkers and cyclists) most sensitive, public transit users moderately sensitive, and car users least sensitive. This gradient reflects how daily mobility conditions intersect with economic capacity, lifestyle, and urban integration experiences. The study contributes to theory by operationalizing behavioral attitudes in a multi-dimensional mobility framework and provides actionable guidance for urban planners. Optimizing transport infrastructure, enhancing neighborhood amenities, and tailoring interventions to specific travel modes can strengthen migrants’ sense of belonging and promote inclusive, sustainable urban development.
Traffic safety research has traditionally emphasized objective environmental factors, with limited attention to how perceived safety interacts with street network characteristics in influencing collision risks. This study developed an integrated framework that combined street view imagery and deep learning to quantify perceived safety at the street-link level in Wuhan, China. We proposed a novel Conditional Autoregressive Zero-Inflated Negative Binomial model that simultaneously accounted for both spatial dependence and zero-inflation in severe collision data. The results showed that street network morphology, including betweenness centrality and diversion ratio, exhibited scale-dependent associations with severe collision frequency. Higher perceived safety was associated with a greater frequency of severe collisions, suggesting a potential safety-risk paradox. In addition, perceived safety significantly moderated the associations between street network characteristics and severe collisions. Specifically, it strengthened the protective association of betweenness centrality at the neighborhood scale, while weakening the protective association of diversion ratio at the urban scale. These findings suggest that perceived safety is not only associated with severe collision frequency, but may also influence how street network characteristics relate to severe collisions. These findings also have important implications for traffic safety planning, particularly in locations where perceived safety may not align with actual severe collision risk, and support the use of visual warning measures, traffic-calming strategies, and surveillance interventions. This study provides new evidence on the complex relationships among perceived safety, street network characteristics, and severe collisions, and offers policy-relevant insights for urban traffic safety interventions.
Understanding travel demand dynamics in rapidly urbanizing environments requires analytical perspectives that transcend traditional disciplinary boundaries. As traffic systems evolve into intricate networks influenced by social interactions, technological innovations, and stochastic behaviors, capturing their non-linear and dynamic nature becomes critical for effective urban management. This paper presents a comprehensive review of ensemble deep learning frameworks and argues that they can serve not only as predictive tools but also as analytical approaches for understanding the complexity of modern mobility systems. Firstly, we synthesize the theoretical underpinnings of ensemble deep learning, highlighting how ensemble strategies (e.g., Bagging, Boosting, Stacking) mirror multifaceted decision-making processes by aggregating diverse model perspectives to handle data heterogeneity. Secondly, we systematically evaluate state-of-the-art deep learning (DL) architectures (including CNNs, GCNs, RNNs and GRUs) and their integration strategies, examining their role in utilizing multi-source big data derived from vehicular sensing and environmental contexts to enhance prediction reliability. Furthermore, this review investigates how ensemble deep learning frameworks enhance system adaptability and generalization by synthesizing the complementary features of multiple models, thereby providing dependable analytical tools for advanced traffic management and planning. Finally, we identify critical challenges and future opportunities, particularly in multimodal data fusion and cross-domain learning, aiming to inform practical strategies for achieving sustainable and resilient intelligent transportation systems (ITS).
Evaluating the supply–demand matching of urban bus systems is essential for optimizing human mobility efficiency and improving residents’ quality of life. However, existing studies primarily rely on static topological metrics derived from complex network theory to assess service capability, often overlooking the critical matching between supply capacity and actual human origin–destination (OD) flows. To address this gap, this study constructs the Supply-Demand Matching and Coverage Index (SDMCI) by integrating supply capability and demand OD flows. Taking Xi’an, China as a case study, we quantify the spatiotemporal variation of bus service matching levels by using proposed SDMCI. Furthermore, we employ the XGBoost-GeoShapley method to explore the spatially heterogeneous contributions of built environment variables to predicted SDMCI. The results demonstrate that the SDMCI in Xi’an exhibits a stable core-periphery spatial structure, while the matching intensity displays significant temporal variations across different time periods. Additionally, the association of the built environment with the SDMCI also shows nonlinear threshold effects and spatial heterogeneity. The SDMCI and analytical framework proposed in this study provide a scientific basis for optimizing allocation of urban public transportation resources, thereby informing transportation planning and decision-making processes.
AI-enabled scoring systems that translate behavioural data into financial consequences are spreading across consumer markets, raising the question of how democratic societies should engage with this emerging form of algorithmic governance (Zuboff, 2019). Pay-how-you-drive (PHYD) motor insurance offers a particularly tangible setting in which to study how citizens experience such systems. The data are continuous, the financial consequences are direct, and a clear transition is now underway from reward-only schemes to schemes that also sanction risky driving. This paper examines that transition empirically. We draw on 19 semi-structured interviews with experienced users of HUK-COBURG’s reward-based Telematik Plus product, all aged 30 to 50, and analyse their reactions to a stimulus describing a sanction-based extension. Following Holland and John (2023), we use the term AI-enabled to refer to the ensemble of machine learning techniques that are at the heart of the HUK-COBURG scoring pipeline. The case is informative because HUK-COBURG’s reward-based Telematik Plus product represents a relatively well-governed implementation in the German market. Qualitative content analysis yields six dimensions of user concern: transparency and complexity, acceptability of risk factors, financial distress, prompting and sanctioning, privacy and data insecurity, and technical issues. We identify what we call a perception-implementation gap. Even in a system designed to exclude socio-demographic discrimination, users perceive themselves to be disadvantaged along precisely those excluded dimensions. We argue that this gap is a structural feature of algorithmic sanctioning more generally, rather than a defect of any single implementation, and we draw out implications for the governance of AI-enabled scoring systems in transportation and beyond.
