Abstract
Urban inequality remains a persistent challenge, often studied through isolated lenses like income or health. We present VulnerabilityMap, an open framework that integrates multi-source data—including census records, human mobility patterns, and social media sentiment—to map the complex landscape of urban disadvantage across the United States. By categorizing challenges into "shocks" (e.g., COVID-19) and hierarchical "stresses" (e.g., social segregation, physiological needs), the framework enables high-accuracy predictive modeling. Our results demonstrate that machine learning can effectively identify the drivers of vulnerability, providing a data-driven foundation for targeted social interventions and policy design.
Introduction
While cities are engines of economic opportunity, they also serve as crucibles of inequality. Millions of residents face structural disparities that limit their access to essential services and social mobility. Historically, research has focused on single dimensions of vulnerability, failing to capture the interconnected "web of vulnerabilities" where factors like environmental hazards, limited mobility, and lack of social capital reinforce one another. VulnerabilityMap addresses this gap by creating a unified pipeline to connect disadvantage features with vulnerability outcomes at the Census Block Group (CBG) level.
Methodology
The VulnerabilityMap framework processes heterogeneous data into a standardized format covering over 328 million people. The methodology is structured around three core components:
- Disadvantage Indices: We constructed four primary indices—Demographic, Neighborhood, Mobility, and Social. These are synthesized into a "Composite Disadvantage Index" using z-score normalization.
- Vulnerability Taxonomy: Drawing inspiration from Maslow’s hierarchy of needs, we categorized long-term "stresses" into Physiological (basic needs), Social (psychological needs), and Self-Actualization (fulfillment needs).
- Predictive Modeling: We utilized Random Forest regression to predict 20 distinct vulnerability outcomes, using feature importance scores to identify the primary drivers of specific risks.
Data Sources
The framework leverages a massive dataset including ACS demographic data, SafeGraph mobility patterns for "experienced segregation," Facebook social connection metrics, and Twitter sentiment indices. Environmental factors like PM2.5 levels and crime statistics are also integrated to provide a holistic view of the neighborhood environment.
Key Insights and Results
Our analysis revealed several critical dynamics in urban inequality. Notably, we observed a "Support Ratio Paradox," where demographically disadvantaged communities often maintain tighter social networks for survival despite fewer resources. During the COVID-19 pandemic, we identified a "Reduce-Rebound" mobility dynamic: while high-status individuals could stay home, disadvantaged populations—often essential workers—were forced to maintain mobility, increasing their exposure risk.
The predictive models achieved high accuracy, with 15 of the 20 vulnerability outcomes reaching an $R^2 > 0.7$. Explainability analysis showed that different "urban assets" require different social or intellectual "entry tickets"; for instance, access to arts and culture is driven by education levels, while access to education facilities is more dependent on social connections.
Conclusion
VulnerabilityMap provides a granular view of urban inequality, showing that geographical proximity does not guarantee similar outcomes. While the current model shows strong predictive power, future work will focus on establishing direct causality and developing a public web interface for policy communication. This framework represents a significant step toward a more holistic, data-driven understanding of urban resilience.
Code and Data: The constructed dataset is available at github.com/LinChen-65/VulnerabilityMap/.
Summary updated