Learning Dependence Relationships of Evacuation Decision Making Factors from Tweets

Individuals react very differently to evacuation orders as their decisions depend on various factors. Identifying key contributing factors and understanding how they affect individuals’ evacuation decisions can help emergency response organizations improve evacuation plans and communication strategies. Conventionally, researchers have studied human evacuation behaviors by conducting post-disaster surveys, which could be costly, be limited by sampling methods, and be dependent on respondent availability resulting in non-timely responses. Social media, becoming an important communication channel during a disaster, can provide alternative data to examine evacuation behavior in near real-time at a relatively low cost. This study explores how social media data can be used to gain insight on human evacuation behavior. We designed a conceptual model, developed a codebook to classify Twitter communications, and employed a Bayesian Network approach to build a model to inductively learn dependence relationships of evacuation decision making factors from tweets. In analyzing tweets during the Lilac Fire in San Diego, CA, the learned Bayesian Network highlighted two key factors, risk perception and received information source, that jointly influenced the individual’s evacuation decision making. This case study also implied that factors related to individual/family situations, evacuation situations, knowledge, and previous experience may not be primary decision-making factors.