Causal Discovery of Urban Design Factors Associated with Public Sentiment Using Geospatial Machine Learning

Abstract

Urban design features such as street greening, enclosure, walkability, and facade complexity are hypothesized to shape how people feel in public space, yet most geospatial machine-learning studies optimize predictive accuracy rather than identify causal relations. This paper develops a causal discovery framework for estimating which urban design factors are associated with public sentiment under explicit identification assumptions. The framework treats street segments or grid cells as units, constructs design exposures from street view imagery and geographic information systems, derives sentiment from geotagged text and visual context, and learns a constrained causal graph over design, contextual, spatial, and sentiment variables. Continuous structure learning with temporal and domain restrictions separates design effects from spatial confounding, residential sorting, and platform selection, while a spatial autoregressive confounder model represents unobserved neighborhood dependence. Effects are standardized through g-computation, and uncertainty is quantified with spatial block bootstrap and graph-stability diagnostics. The contribution is not a new sentiment classifier but an end-to-end causal estimand and inference design that links multimodal geospatial measurement to intervention-oriented questions in urban design. The analysis shows why predictive importance, partial correlations, and ordinary spatial regression can misstate design effects when sentiment is selectively reported and design features co-vary at multiple scales. It also specifies boundary conditions for interpretation, including positivity, interference, measurement invariance, and graph faithfulness. The framework provides a reproducible basis for evaluating design scenarios, prioritizing street-level interventions, and reporting causal claims with appropriate spatial and measurement uncertainty.