Fairness-regularized geographically weighted regression for urban accessibility analysis

The increasing use of predictive models in socially sensitive domains has intensified concerns regarding the reproduction of historical biases affecting vulnerable population groups. Although fairness-aware machine learning has developed numerous methodologies to mitigate unfair predictive behaviour, these approaches have been largely concentrated on non-spatial applications such as finance, healthcare, or recommendation systems. In contrast, urban-planning studies have extensively documented socio-spatial disparities related to accessibility to public services and urban opportunities. Consequently, predictive models trained on urban data reflecting such inequalities may perpetuate or intensify existing accessibility gaps.

Motivated by real urban-accessibility datasets exhibiting significant socioeconomic disparities, this paper introduces a fairness-regularized extension of geographically weighted regression (GWR). Specifically, the proposed methodology incorporates explicit fairness constraints into the classical locally weighted least-squares formulation in order to regulate predictive disparities between sensitive and non-sensitive population groups. The resulting constrained regression problem can be efficiently solved using standard quadratic-optimization routines. 

The proposed methodology is illustrated through two urban-accessibility case studies in the city of Seville (Spain) involving Urban Green Spaces and Healthcare Facilities datasets. Numerical experiments reveal a clear trade-off between predictive accuracy and fairness, while also highlighting the spatial redistributive effects induced by fairness constraints.