Research

An end-to-end pipeline turns mobile device location data into individual trip rosters via home/work identification, trip detection, mode imputation, population weighting, and validation. Rosters aggregate into multi-modal OD matrices at flexible spatiotemporal resolutions.

Can mobile-device OD data reproduce road-level traffic flow? We map mobile OD onto road networks via dynamic traffic assignment across 35 metros in Mexico, Colombia, Indonesia, and India, finding persistent underperformance in low-penetration regions.

Forecasted trip itineraries integrate with micro (Vissim, SUMO), meso (DTALite), and agent-based simulators (MATSim), enabling fine-grained citywide traffic simulation driven entirely by mobile location data and behavioral patterns.

Multi-ATGCN, a Multi-graph Multi-head Adaptive Temporal Graph Convolutional Network for citywide multi-step OD flow prediction, achieves consistent accuracy gains over state-of-the-art baselines on two real-world datasets.

A hierarchical activity-based framework jointly predicts the activity, time, and location of each device's next trip. Loss functions borrowed from semantic segmentation handle severe class imbalance across 18,000+ residents at county scale.

Treating nationwide CBG-level mobile-device trip flows as a travel-demand proxy, we benchmark 48 explainable ML models and interpretation techniques, documenting strong nonlinearities, threshold effects, and feature interactions.

Computer vision maps vehicle-level emissions from urban traffic video. A 2.2M-image dataset classifies 4,923 car models, and per-vehicle pollutants are estimated via a modified COPERT factor incorporating velocity and acceleration extracted through vehicle tracking.

A ubiquitous data-driven framework for city-scale traffic emission estimation that fuses camera videos with mobile phone data. Inferring directional signal timing shows PM and NOx emissions rise 40 to 60 percent on average when stop-and-go cycles are properly modeled.

Quantifying city-wide effects of 30 km/h speed limits in Milan via 3.4 million driving trips. Aggregate emission changes are modest (0.6 to 2.7 percent) but vary substantially across space and time, peaking during cross-zone trips at rush hour.

A Digital Twin platform for road traffic emission nowcasting and forecasting. Integrating online repositories with IoT sensors enables alternative-policy scenario evaluation, demonstrated for Kista, Stockholm via 3D interactive visualization.

Discrete-event simulation quantifies how battery capacity affects shared electric vehicle adoption. Constrained capacity hurts both user satisfaction and vehicle utilization, while faster charging, larger range, and higher vehicle-to-trip ratios mitigate the impact.

A bi-level optimization framework for dynamic wireless charging of battery electric buses, jointly handling strategic facility deployment and battery sizing alongside tactical charging scheduling under time-of-use electricity tariffs.

Five million transactions from EVCARD reveal that users avoid older, smaller, and pricier vehicles. State of charge dominates the choice; users tend to be "greedy," consistently selecting max-SoC vehicles even when their actual trips are short.

Tracking US pandemic-era mobility via 150 million monthly active mobile devices, capturing trips per person, person-miles traveled, and share of residents staying home. Updated daily on a public platform.

Twenty years of Chicago transit ridership analyzed via Bayesian structural time series to isolate the pandemic's causal impact while controlling for socioeconomic confounders. Sharpest declines hit commercial areas and white, educated, high-income neighborhoods.

Chicago bikesharing across the pandemic, benchmarked against transit, driving, and walking. Bikesharing proved the most resilient mode, yet stark socioeconomic inequities surfaced as high-income station catchments saw both the steepest gains and declines.

Mobile-device data reveal a spontaneous mobility reduction that preceded government mandates, plus a "floor" effect once stay-at-home orders took hold. State-level policies accounted for only roughly five percent of the total mobility decline.

A comprehensive survey of multi-modal urban transport network resilience covering modeling, evaluation, and optimization across road, transit, and shared mobility systems, with an outlook on next-generation infrastructure.

Fusing social media posts, precipitation records, and traffic flow data to automatically detect urban road flooding in Shenzhen. The system achieves a 68 to 90 percent detection rate with only 1.5 to 2 percent false alarms.

Comparing passively collected mobile location data with actively reported Waze incidents for monitoring extreme weather impact on traffic. Active reports underestimate severity, since fewer users venture out as conditions worsen and become silent sensors.

Quantifying nationwide changes in mobility inflow at the pandemic's onset and modeling its time-varying relationship with infection rates, with the dynamic positive association substantially stronger in partially reopened regions.

Linking COVID-19 outcomes to demographics, occupation, and partisanship reveals a structural inequality: social distancing operates as a "privilege" of advantaged groups, while disadvantaged communities bear disproportionate case and death burdens.

Structural equation modeling on 4.4 million POI geo-tracking records shows that mobility mediates COVID-19 infection rates but not case-fatality ratios, with substantial variation across county-level racial and ethnic compositions.

Mediation analysis demonstrates that stated vaccine hesitancy alone cannot fully explain US vaccination disparities, challenging the dominant narrative and pointing to structural access barriers as a complementary driver of inequity.

Two years of US county data on vaccination, mobility, and outcomes show that vaccine effectiveness against case rates diminished during the Omicron surge, while protection against case-fatality persisted. An NIH-funded agent-based extension is underway.

As technical lead of incenTrip, the first nationwide incentive-based travel demand management app, we design personalized dynamic incentives for mode switching, carpooling, congestion avoidance, and off-peak travel.

Four million coupons analyzed via extended Cox proportional hazards models. Coupon effectiveness peaks early and declines after roughly 130 days, so front-loading new-user incentives maximizes short-term revenue but raises fairness and privacy concerns.

Analyzing 2.7 million food delivery orders in Dubai to explain why customers choose distant restaurants over closer alternatives. Explainable ML reveals that delivery fees, cuisine type, restaurant ratings, and neighborhood socioeconomics shape whether orders stay local.

Five million parking-related Google Maps reviews across 1.1 million US POIs are classified with BERT to model socio-spatial sentiment. Restaurants register the most negative scores, while denser and lower-income urban areas show consistently negative perceptions.

A Llama 3 model with Low-Rank Adaptation fine-tuning measures public accessibility sentiment from nationwide Google Maps reviews. Older and more educated areas express more negativity, while disability prevalence alone shows no clear association with sentiment.

Shanghai dockless bikesharing evaluated for bike-and-ride (BnR) integration with metro across four metrics. Land-use mix correlates with BnR trips only beyond a 1.5 km buffer, and larger operators generate more BnR trips but no higher per-bike utilization.

Station-level booking and turnover analysis for carsharing optimization shows that operators should target underserved areas. Carsharing competes best 1.2 to 2.4 km from bus stops, and geographically differentiated quotas help local authorities manage fleets.

Comparing e-scooter sharing with docked bikesharing usage via generalized additive mixed models. Both modes succeed in dense, young, higher-income areas, yet a one-sided competitive relationship has emerged with e-scooters surpassing bikes in adoption.