A BOTTOM-UP AND DATA-DRIVEN APPROACH FOR UNDERSTANDING RISK COMMUNICATION AND STAKEHOLDER ENGAGEMENT TO INFORM EVACUATION AND PROTECTIVE DECISIONS IN HAZARD-PRONE COMMUNITIES .

Abstract

Effective risk perception and communication are essential components in enhancing disaster resilience and ensuring timely protective actions. Decision-making during disasters is a complex process influenced by a variety of factors, including individuals' perceptions of risk, the trust they place in sources of information, and their communication patterns. However, a systematic review shows that literature has largely emphasized top-down (from authority to individual) communication strategies and quantitative assessments, with limited attention to bottom-up, household-level understanding of how individuals perceive, interpret, and act upon disaster information. Furthermore, there is a lack of integrated frameworks that connect in-person decision-making processes with emerging communication dynamics in online social media environments. Addressing these gaps, this study adopts a combined qualitative and data-driven approach to examine disaster risk perception, communication and user interaction patterns. Part I of the study employs a bottom-up qualitative methodology to investigate how residents in Oklahoma perceived, communicated, and responded to tornado warnings during two major events in April and November 2024. Based on 100 in-depth interviews with both directly affected individuals and those exposed to warnings, the analysis identifies seven key themes: (i) risk perception, (ii) information sources, (iii) trust in information, (iv) risk communication, (v) resource sharing, (vi) decision points, and (vii) aftermath. The qualitative findings provide critical insights into how individuals perceive, communicate, and respond to tornado risk, while also revealing the growing importance of informal and digitally mediated communication channels. Participants consistently reported limitations in official warning systems, including delayed alerts and insufficient communication of storm severity. As a result, individuals increasingly relied on peer-to-peer communication through social media, text messaging, and local networks to verify information, share updates, and coordinate responses. This shift highlights the central role of decentralized information flows and trust-worthysocial networks in shaping protective decision-making. The reliance on social media and interpersonal networks, particularly in situations where institutional communication is perceived as inadequate, suggests that online platforms serve as critical extensions of real-world communication processes. Additionally, variations in response behavior shaped by trust, prior experience, and household context indicate that communication patterns are heterogeneous and influenced by user characteristics. Building on the qualitative findings, Part II of the studyadopts a data-driven framework to analyze communication and interaction patterns on Twitter (X). Supervised machine learning models are used to classify users based on multi-dimensional features, including user engagement metrics, sentiment, temporal attributes, and topic information, along with environmental variables (e.g., weather conditions).Decision Tree Model achieved the highest performance (accuracy > 92%). To capture relational dynamics, a deep neural network–based pairwise model is developed, achieving an accuracy of approximately 78% in predicting user interactions. The results reveal that user interactions are highly structured, with strong homophily observed among individual users, who dominate communication and act as central connectors within the network. Cross-group interactions indicate their role in bridging different user categories (bot, individuals, media, agency, others). Community detection further identifies both highly centralized and heterogeneous clusters, reflecting variations in interaction intensity and behavioral patterns. Overall, the findings demonstrate the effectiveness of the proposed data-driven models in capturing both user classification and interaction dynamics in disaster-related communication. Overall, the bottom-up qualitative approach provides critical depth by capturing lived experiences of affected individuals, uncovering how trust, information sharing, and risk perception dynamics shape decision-making under uncertainty. Building on these insights, this study contributes a unified framework that integrates qualitative understanding with data-driven modeling to enable a rule-based and well-structured organization of social media users based on their behavioral roles and interaction patterns. This approach supports more effective risk communication strategies, facilitates monitoring of information diffusion, and enhances the design of targeted interventions. Ultimately, these findings contribute to improved strategic disaster resilience planning. Future research will extend this framework by incorporating real-time data streams, multi-hazard contexts, and advanced data-driven techniques to further strengthen predictive capabilities and decision support systems.