Effective transit service planning requires data-driven decision-making, but traditional methods often rely on subjective analysis of complex flow visualizations. This presentation introduces the Foursquare Integrated Transportation Planning (Foursquare ITP) Corridor Identification model, a tool designed to identify high-demand corridors for high-frequency transit services using origin-destination flow data and an optimization algorithm. By pinpointing travel corridors with the highest ridership potential, this tool provides planners with data-backed insights to design efficient and effective transit networks. The Corridor Identification model processes origin-destination data—such as agency surveys, Location-Based Services (LBS) data, and regional travel models—through an optimization algorithm to identify corridors with maximum demand.

The Corridor Identification model is adaptable, allowing planners to work with different datasets, time periods, and trip attributes to reflect varied travel patterns. For example, the tool consumed both synthetic trip data from Replica and VIA Metropolitan Transit’s on-board survey dataset to identify key corridors for San Antonio’s VIA Best Bus Network Redesign, revealing areas where high transfer rates could be alleviated through route adjustments. Similarly, for WMATA’s Better Bus Network Redesign in Washington, D.C., the model helped generate demand-driven corridors that guided service decisions, confirming the feasibility of proposed network changes. Additionally, the model has been used in the TxDOT Statewide Active Transportation Priority Network Development project, identifying potential trail corridors across Texas, based on the highest levels of active transportation trips. This effort helps ensure that the state’s active transportation network aligns with real-world travel patterns, promoting more sustainable transportation options statewide. These case studies highlight how the tool not only generates corridor candidates but also verifies and justifies planning decisions with quantitative evidence.

The development of the Corridor Identification model bridges academia and industry. Originally introduced in a PhD dissertation as a mixed integer linear program, the algorithm was further enhanced by Foursquare ITP to serve commercial uses in the transportation planning industry. The model is thus an example of how academic research can drive innovation and meet the practical needs of transit agencies, facilitating collaborative solutions between research and practice.

This presentation will showcase how the Corridor Identification model enhances the planning process by connecting data to decision-making. Attendees will see how this tool can be adapted to various datasets and cities, offering a practical, open-source solution to improve transit planning. The model’s ability to identify demand-driven corridors ensures that service redesigns better meet the needs of passengers, improving both efficiency and passenger experience.

AEEMAT is an interactive modeling tool for predictive analytics of fuel consumption by traditional vehicles and the resulting air pollutant emissions, as well as electricity consumption by EVs and the associated emissions from power plants, under various transportation planning and investment policies in the Atlanta Metropolitan Area. AEEMAT integrates ARC's ABM with a sophisticated machine learning model, trained using the MOVES inventory and Cambium database, enabling accurate assessments of energy consumption and emissions from both vehicles and the power grid. The user-friendly interface designed for AEEMAT will enable comprehensive analysis of planning and policy scenarios, aligned with ARC's 2050 MTP.   

Housing prices in urban areas are influenced by a complex interplay of transportation accessibility, infrastructure quality, and neighborhood demographics. Studies have shown that accessibility to public transit and road networks can enhance property values. For example, (Martínez & Viegas, 2009) found that proximity to metro stations increases housing prices, while (Debrezion et al., 2011) reported that homes near Dutch rail stations are approximately 25% more expensive than those 15 km away. However, not all transportation infrastructure has a positive effect. While access to transit can be an amenity, proximity to highways may act as a disamenity due to noise, pollution, and safety concerns (Mathur, 2008). In some cases, housing prices increase with distance from transit facilities, suggesting that certain transportation infrastructure may be perceived negatively (Acuña, 2023). Thus, local or city specific studies that account for the residents’ preference, current market scenario and transportation availability are necessary to understand the nature of the relationship for that city/region.

In this paper, we investigate property prices in the Denver metro area, one of the fastest-growing metropolitan regions in the U.S. post-pandemic. Rising demand has put increasing pressure on Denver’s housing market, threatening affordability and access to homeownership. Our study presents a predictive model for housing prices in Denver using 2020 market value data, integrating key variables related to transit accessibility, built environment characteristics, and neighborhood sociodemographics to enhance the understanding of housing market dynamics. Property values, including lot sizes and building types, are sourced from the local assessment authority. Transit accessibility is assessed using data from the Regional Transportation District (RTD) of Denver, capturing the distribution of bus and light rail networks, station locations, and service frequencies. Built environment characteristics, such as road connectivity and pedestrian infrastructure quality, are obtained from the Environmental Protection Agency’s (EPA) Smart Location Database. Demographic and economic factors, including household income and population density, are sourced from the U.S. Census Bureau’s American Community Survey (ACS). To refine the analysis, this study develops latent variables to assess neighborhood quality and multimodal connectivity across all census block groups (CBGs) in the study area. Neighborhood quality reflects factors such as green space availability, job-residence balance, and network density, while multimodal connectivity considers sidewalk and bike infrastructure, transit stop density, and public transit accessibility. Building on these indicators, we use machine learning algorithms to predict property prices, uncovering complex relationships and improving valuation accuracy.

The findings offer insights into the factors shaping housing prices in a growing metro area like Denver, helping planners and policymakers develop more reasonable housing policies. As cities expand and interest in transit-oriented development rises, understanding housing price dynamics is crucial for shaping effective land-use policies and housing affordability strategies. A data-driven approach can help mitigate unintended consequences such as gentrification and displacement while ensuring transit investments benefit diverse communities.

Abstract Background

Conventional Travel Demand Models (TDM) produce forecasts with aggregated inputs and fixed response functions, limiting fidelity for short-term conditions when both demand and network characteristics change. While time series machine learning (ML) methods perform well under stationary network conditions, their accuracy declines sharply when capacity changes, such as when widenings or closures occur. We propose a hybrid deep-learning framework that forecasts link-level traffic volumes by fusing two complementary evidence sources: (i) historical observations capturing real temporal behavior, and (ii) CUBE simulations encoding counterfactual network changes (e.g., capacity adjustments). The model couples a temporal graph transformer with a spatial graph transformer and learns data-driven fusion weights to produce period-specific volume forecasts. We chose graph transformers because this single model architecture enables unified spatial and temporal modeling, aligns historical and simulation data for joint inference, and scales effectively across network sizes and forecasting horizons. While the primary forecast target is traffic volume, the entire framework is developed within a congestion pricing context, where understanding revenue implications is inherently linked to traffic behavior and, thus, revenue is additionally forecasted.

Description of Abstract

We represent the study area network as a directed graph with nodes (zones/intersections) and edges (A–B connections). Edge features include previous and current capacities, lagged volumes and revenues, exogenous indicators (employment, unemployment, work-from-home percentage), and calendar variables (day, month, year, day-of-week).

Two encoders operate on aligned but distinct views of the same topology: The temporal graph transformer takes in historical edge sequences to model autoregressive and seasonal effects, while the spatial graph transformer takes in CUBE runs where capacity deltas perturb the network to learn structural sensitivity. A weight (gating) decoder performs adaptive fusion via learned attention over the two streams and outputs per edge.

Our methodology comprises five stages: (1) Generate training samples by applying capacity changes at both the corridor level and the global network level using CUBE simulations, and construct topology‐consistent graphs from these outputs while aligning historical transactions/gantries to corresponding A–B links; (2) generate and normalize features across both datasets; (3) jointly train temporal and spatial encoders with a fusion decoder under a unified loss; (4) optionally apply a corrective signal to the fusion decoder during training to guide weight adaptation; and (5) evaluate performance using temporal holdout and scenario‐based holdout with unseen capacity changes.

Validated using real-world data from the I-66 Inside the Beltway (ItB) Express Lanes corridor in Northern Virginia, the model produced an average Mean Absolute Percentage Error (MAPE) in the range of 8–11% for temporal forecasts across all gantry links, capturing both short-term fluctuations and seasonal patterns. To assess performance under structural changes, we generated predictions for 64 CUBE-derived capacity‐change scenario graphs representing varied magnitudes and spatial distributions of link modifications. Across these scenarios, the average difference between model predicted and CUBE predicted values was 3.4% MAPE, indicating strong robustness to network perturbations, with the highest error for a link observed at 10.03% MAPE.

Statement on Why Abstract is Noteworthy

The architecture explicitly handles joint demand and network shifts, enabling robust short-term forecasting and credible what-if analysis under capacity changes. This method scales to larger, multi-facility networks and as future work can be extended to directly incorporate toll policy levers, enabling downstream optimization objectives such as maximizing throughput or revenue, making it a practical decision-making support tool for planners operating under evolving conditions.

Project is Ongoing