Reducing AI Model Biases with a Bilevel Learning Framework: A Case Study of Leveraging Twitter Data for Damage Estimation
Weishan Bai, Xinyue Ye, Yiqun Xie, Shannon Van Zandt, Xiao Huang, Debalina Sengupta · Annals of the American Association of Geographers · 2025
This study aims to improve disaster risk mitigation by integrating fairness into artificial intelligence (AI) models, specifically addressing spatial biases that can lead to unequal resource allocation during disasters. Our objective is to reduce spatial biases in disaster impact prediction models via a bilevel learning framework, enhancing both accuracy and fairness. To achieve this, we leverage information exchanges in social networks during the disaster period to estimate the economic loss caused by the disaster. Considering that accessible data are spatially biased in quantity and quality, this study applies a fairness framework within a deep neural network to mitigate spatial bias in predictions. We analyze Twitter data and the Federal Emergency Management Agency’s Real Property Damage Amount data to assess and predict the economic impact of Hurricane Harvey in Texas at the census block level. By integrating a bilevel learning framework within a deep neural network model, we specifically target and reduce spatial biases. Our results demonstrate that this framework not only improves prediction accuracy, particularly in areas with low population density, but also ensures more disaster response strategies with improved fairness. This study provides a novel contribution to the field by showcasing how AI models can be adapted to ensure fairness in disaster management, offering valuable insights for future AI-driven disaster assessment and response systems. Our results also promote the need for fairness in AI-driven disaster response to prevent unequal resource allocation, as evidenced in recent case studies on Hurricane Harvey.