Adaptive Cross-City Disaster Sentiment Learning

Social-media models often reproduce urban bias because data-rich cities dominate online discussion while rural and disadvantaged communities remain underrepresented. This project develops a context-sensitive framework for more equitable disaster sentiment understanding.

Technical innovation

The two-layer architecture combines socioeconomic city-similarity modeling with multimodal text-mobility fusion. Similarity-weighted adaptation transfers knowledge to data-sparse cities while preserving local sentiment signals, and mobility embeddings connect online expression to physical behavior.

Practical insight

Applied to the January 2025 Southern California wildfires, the framework reveals geographically diverse sentiment patterns in places with overlapping fire exposure or delayed emergency response. More calibrated estimates can help decision-makers see needs that urban-centered models overlook.

Related publication

Bridging the Urban Divide: Adaptive Cross-City Learning for Disaster Sentiment Understanding