Integrating information from heterogeneous networks on social media to predict post-disaster returning behavior

With the recent rise in frequency and intensity of severe disasters, increasing the resilience of urban systems is an urgent challenge. Local governments need to accurately predict the demand of public services such as gas, water and power after disasters in order to effectively allocate resources for recovery, and also to prevent secondary disasters. In practice, post-disaster surveys and interviews have been used to grasp the returning dynamics of evacuees. Here we propose a novel method for integrating information obtained from heterogeneous networks on social media that can be used to improve the predictability of returning behavior of evacuees after severe disasters. We validate our proposed method using a large Twitter dataset from Hurricane Sandy and verify that the model is capable of predicting 86% of the returning behavior of evacuees correctly, increasing the performance by 21% compared to conventional methods. Results imply that tweets posted near the individual's residential location and posts of online connected peers have significant marginal effect on predicting individuals’ returning behavior. The proposed method and gained insights could help decision makers in various agencies to predict the demand of public services in affected regions immediately after the disaster, enabling them to plan effective and efficient infrastructure repairing strategies given time and budget constraints.