A Framework to Model Human Behavior at Large Scale during Natural Disasters

Abstract

Natural disasters affect millions of people every year. Understanding human behavior is critical to improve both emergency planning and prevention. However, emergency responders typically struggle to gain access to timely, fine-grained models of human behavior during disasters. In this paper, we propose a novel framework to analyze behavioral changes during disasters using Call Detail Records (CDRs) from a telecommunications company. CDR datasets are collections of spatio-temporal traces that can characterize individual mobility and social network behaviors at very fine scales. The proposed framework exploits the granular behavioral models to evaluate the similarities and differences between the pre and post-disaster patterns. The framework consists of three steps: data pre-processing, behavioral baseline modeling and disaster analytics. The data pre-processing step uses data mining techniques to extract individual mobility traces and individual social network features from the CDR data. The behavioral baseline modeling step computes the baseline models that characterize normal mobility and social network behaviors in non-disaster scenarios. This step uses n-th order Markov Chain models to approximate mobility patterns from the CDR spatio-temporal data. Finally, the disaster analytics step allows for the statistical analysis of behavioral changes during disasters by comparing the real behaviors observed during a disaster with the behaviors that would have been expected under normal circumstances (extracted using the baseline models). We use the framework to analyze Rwanda's 2012 floods and show that disasters tend to disrupt both mobility patterns and communication behaviors while recovery times can take several weeks.