Harnessing Generative Models for Equity in Transportation: A Survey

Abstract: Generative models have demonstrated remarkable potential in addressing data equity challenges in transportation by generating realistic, high-quality data from limited samples. These models, including Large Language Models (LLMs), Variational Auto-encoders (VAEs), Generative Adversarial Networks (GANs), and Diffusion Models (DMs), offer innovative solutions to augment datasets, generate diverse scenarios, and promote more inclusive and equitable transportation systems.

This survey explores the use of generative models to advance transportation equity by addressing social and data-related challenges across various stages of the transportation data pipeline, including policy-making, data collection, data processing, and data dissemination. By tackling these challenges, generative models can help create more representative and inclusive datasets, enabling better-informed and equitable decision-making processes in transportation.

We provide an introduction to the key concepts and architectures of LLMs, VAEs, GANs, and DMs, highlighting their specific roles in data augmentation and scenario generation for transportation equity. For example, LLMs can generate synthetic transportation-related text data, such as user feedback and policy documents, to improve the representativeness of marginalized groups. VAEs can generate synthetic traffic flow patterns, ridership data, and user profiles while preserving the statistical properties of the original data. GANs can generate realistic road network layouts, land use patterns, and mobility traces with high fidelity and diversity. DMs can generate synthetic traffic camera footage, lidar point clouds, and satellite imagery, enabling more diverse and inclusive data collection and analysis.

Furthermore, we discuss our insights for future research directions, providing a fresh perspective for agencies and researchers to leverage generative AI techniques in creating more inclusive and equitable transportation systems. These directions include developing evaluation metrics for assessing the fairness and representativeness of generated data, integrating domain knowledge into generative models, exploring the use of generative models for participatory planning and community engagement, addressing ethical and privacy concerns.

By harnessing the power of generative models, transportation agencies and researchers can create more representative and equitable transportation systems that better serve the needs of all individuals and communities. This survey aims to provide a comprehensive overview of the opportunities, challenges, and future directions in leveraging generative AI techniques to promote transportation equity, paving the way for more inclusive and sustainable transportation solutions.