Electric vehicle timeshare rentals (EVTRs) have been recognized as promising solutions to growingly severe problems of traffic congestion, air pollution, and insufficient parking spaces. This study aims to explore the factors that affect the adoption of EVTRs. To achieve the research objective, the household survey is conducted to obtain the travelers’ attitudes towards their travel. Ten latent attitudinal factors are extracted based on the technology acceptance model (TAM) and the theory of planned behavior (TPB). The multi-index and multi-cause (MIMIC) method simultaneously estimates the correlations between the attitudinal factors. Two logit models with attitudinal factors or not are constructed to estimate the quantitative relationship between various factors and EVTR adoption. The results show that the accuracy of the mixed logit model with latent attitude variables is better than the binary logit model without attitude latent variables. This indicates that attitude latent factors could be well matched with the traveler’s travel behavior and could better reflect travelers’ travel demand. Perceived comfort, perceived efficient, subjective evaluation, use preference, and use willingness significantly impact EVTR use frequency. The inconvenience in travelers, rental stations, shared vehicles, and use modes have significant negative impacts. Finally, social pressure has no significant impact. Findings provide valuable insights regarding the efficient planning of the EVTR system and allow decision-makers to develop scientific and practical measures of EVTRs.

E-bikes have gained significant popularity in China from their convenience and eco-friendliness, but the lack of helmet use among riders poses serious head injury risks. This study evaluated the effectiveness of safety campaigns and educational initiatives in promoting helmet adoption among e-bike riders in Guangdong Province, China, across three perceived effectiveness levels: ineffective, moderately effective, and effective. Data from a questionnaire survey were analyzed using an optimized eXtreme Gradient Boosting (XGBoost) model, with hyperparameters tuned via Optuna and performance validated through 10-fold cross-validation, achieving 84.6% accuracy. Class imbalance was addressed using the synthetic minority over-sampling technique, yielding macro-averaged precision, recall, and F1-scores of approximately 0.82, indicating balanced predictive performance. Model interpretability with SHapley Additive exPlanations (SHAP) identified safety education frequency, enforcement of helmet use, riding experience, ideal education locations, attitudes toward helmet use, and age group as the most influential predictors of helmet adoption effectiveness. Safety education frequency contributed most at the ineffective (25.36%) and effective (18.65%) levels, while enforcement was most influential at the moderately effective level (19.11%), with persuasion-plus-fines outperforming education-only strategies. Partial dependence plots (PDPs) and individual conditional expectation (ICE) plots showed that monthly or quarterly education substantially improves perceived effectiveness, whereas infrequent or no education increases perceived ineffectiveness. Younger riders (under 25 years) and those with negative helmet attitudes were more likely to perceive initiatives as ineffective. Findings highlight actionable recommendations: institutionalizing frequent safety education in high-impact locations, integrating enforcement with education, and tailoring interventions to demographic and attitudinal profiles, offering policymakers evidence-based approaches to enhance helmet adoption and rider safety.

This study aims to investigate the heterogeneity in bus-involved crashes by identifying association rules and contributory patterns of road facilities across different crash types and severities. The goal is to support more targeted and context-specific safety interventions. A total of 14,560 bus-involved crashes that occurred in London, UK, from 2010 to 2019 were analyzed. Each crash record includes attributes related to the driver, vehicle, environment, and road infrastructure. To address crash heterogeneity, Latent Class Clustering (LCC) was used to classify crashes into distinct clusters based on uncontrollable factors such as driver characteristics and environmental conditions. Subsequently, Association Rule Mining (ARM) was applied to identify contributory chains and association patterns of controllable road infrastructure factors within each cluster and severity level. The top ten rules for both slight and serious crashes were extracted in each group. The findings reveal that typical features and association rules vary significantly across crash clusters and severity levels. On one hand, a single associative chain can result in varying severity levels across different crash clusters. For example, dual carriageways and daylight are contributing factors to both slight and severe crashes. On the other hand, associative chains leading to the same severity level may differ among crash clusters. For instance, yield or uncontrolled junctions, zebra crossings, and daylight are key factors in slight crashes within Cluster 3, but not in other clusters. The proposed framework, which integrates LCC and ARM, enables a more nuanced analysis of bus-involved crash data by separating classification and association rules analysis based on controllability. The results provide actionable insights into how specific road infrastructure features contribute to different types and severities of crashes. This can inform the design of targeted safety measures and infrastructure improvements to reduce crash risks in urban bus systems. Moreover, the framework is adaptable to other transportation modes and crash datasets.

Coordinated ramp metering (CRM) is an effective strategy for mitigating congestion on urban expressways. It utilizes traffic flow measurements from the expressway network to regulate multi-ramp signals for optimal system performance. Many CRM approaches adopt a reactive control framework in which control actions are initiated only after congestion has been detected. This reactive mechanism may delay control responses and thereby reduce the effectiveness of interventions under rapidly changing traffic conditions. To address this, the paper presents a predictive CRM method integrated with a dynamic spatial–temporal graph convolutional prediction model (CRM-DSTGC). Utilizing the individual trajectory information extracted from automatic vehicle identification (AVI) data, we first define the road segment unit, road segment group, and bottleneck. A deep learning prediction model is then developed based on graph convolution to capture the dynamic spatial–temporal dependencies among road segments and the interactions between related traffic variables. Based on the observed and predicted traffic states, the ramp controllers determine the green durations for the on-ramps using the multi-agent deep reinforcement learning algorithm MADDPG. Finally, the proposed method is evaluated in a SUMO-based simulation of a real-world urban expressway. In traffic state prediction experiments, the proposed DSTGC model achieves Mean Absolute Percentage Error (MAPE) values of 6.966% (3-step), 8.726% (6-step), and 9.871% (12-step), thereby outperforming baseline methods. This enables short-term traffic-state prediction and provides predictive traffic information for CRM. In CRM simulations, the CRM-DSTGC method reduces the mainline congestion duration by 58.33% compared with the prediction-free CRM strategy. The proposed method provides a potential approach for improving CRM in urban expressway networks.