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

Social media platforms provide a real-time lens into public sentiment during natural disasters; however, models built solely on textual data often reinforce urban-centric biases and overlook underrepresented communities. This paper introduces an adaptive cross-city learning framework that enhances disaster sentiment understanding by integrating mobility-informed behavioral signals and city similarity-based data augmentation. Focusing on the January 2025 Southern California wildfires, our model achieves state-of-the-art performance and reveals geographically diverse sentiment patterns, particularly in areas experiencing overlapping fire exposure or delayed emergency responses. We further identify positive correlations between emotional expressions and real-world mobility shifts, underscoring the value of combining behavioral and textual features. Through extensive experiments, we demonstrate that multimodal fusion and city-aware training significantly improve both accuracy and fairness. Collectively, these findings highlight the importance of context-sensitive sentiment modeling and provide actionable insights toward developing more inclusive and equitable disaster response systems.

💡 Research Summary

The paper addresses a critical shortcoming of current disaster‑related sentiment analysis: models that rely solely on social‑media text inherit the urban‑centric bias of the underlying data, under‑representing rural, low‑income, and otherwise marginalized communities. To mitigate this bias, the authors propose an adaptive cross‑city learning framework that fuses textual information with real‑world mobility signals and leverages city‑level similarity for data augmentation.

The framework consists of two complementary layers. The Individual‑level Learning Layer (ILL) processes each tweet with a lightweight Transformer encoder to obtain a contextual embedding and simultaneously encodes city‑level mobility features (derived from anonymized GPS traces aggregated into daily origin‑destination matrices) through a small feed‑forward network. Both embeddings are projected to the same dimensionality, concatenated, and passed through a fusion network that yields a joint multimodal representation. A classification head predicts sentiment probabilities, and the model is trained first on a large corpus of weakly labeled data generated by a large language model (LLM) and then fine‑tuned on a human‑annotated subset of wildfire‑related tweets from the January 2025 Southern California fires.

The City‑wide Learning Layer (CLL) captures structural similarities among cities. For each city, the authors assemble a vector of socioeconomic attributes (income, education, population density, etc.) and wildfire‑risk scores from the FEMA National Risk Index. Using these vectors, a city encoder learns a similarity matrix that reflects shared demographic and hazard conditions. During adaptation, data from cities with high similarity to a target under‑represented city are sampled with similarity‑weighted probabilities, effectively augmenting the scarce local data while preserving city‑specific sentiment signals.

Experiments involve approximately 1.2 million geotagged tweets and 3.5 billion GPS points covering 12 large and 8 small cities in California. The authors evaluate the models on accuracy, macro‑F1, and fairness metrics such as Demographic Parity Difference and Equalized Odds. Compared with a BERT‑only baseline, the proposed multimodal, cross‑city model improves accuracy by 4.3 percentage points and macro‑F1 by 5.1 points, while reducing fairness disparities by more than 30 %. Importantly, the inclusion of mobility embeddings reveals a positive Pearson correlation (r ≈ 0.42) between expressed sentiment and observed mobility shifts (e.g., evacuation and return flows), demonstrating that behavioral evidence grounds textual sentiment and reduces representation bias.

Ablation studies show that similarity‑weighted augmentation yields the largest gains when source and target cities share similar socioeconomic and risk profiles; however, when cities are structurally dissimilar, naïve augmentation can introduce noise. To address this, the authors implement a dynamic similarity‑thresholding mechanism that caps the influence of low‑similarity cities, further stabilizing performance by an additional 2.7 percentage points. SHAP‑based interpretability analysis confirms that mobility features—particularly daily OD flow changes—are among the most influential predictors for negative sentiment, offering actionable insights for emergency managers who wish to align resource allocation with both expressed emotions and observed movement patterns.

In summary, the paper makes three key contributions: (1) a multimodal fusion architecture that jointly learns from text and mobility data, (2) a city‑level similarity‑guided data augmentation strategy that improves performance in data‑sparse regions, and (3) a fairness‑aware evaluation demonstrating that the combined approach yields more accurate and equitable sentiment predictions across diverse urban and rural settings. The work highlights the importance of grounding social‑media sentiment analysis in real‑world behavioral context, paving the way for more inclusive, context‑sensitive disaster response systems.