Interpretable boosting tree ensemble method for multisource building fire loss prediction

Abstract

Building fires may cause enormous property loss. Disaster relief organizations and post-disaster recovery efforts benefit from the accurate and interpretable prediction of property loss. To solve this problem, we propose a novel interpretable boosting tree ensemble method (IBTEM), which is of practical significance for providing decision support for dispatching aid and mobilizing recovery resources. First, we fuse multisource datasets including National Fire Incident Reporting System (NFIRS) dataset and National Oceanic and Atmospheric Administration (NOAA) dataset from 2012 to 2016. Second, we construct four variable scenario subsets to select related variables for building fire loss. Third, we adopt Winsorization, logarithmic transformation, recursive feature elimination and weighted voting strategies to create an ensemble of boosting trees. Fourth, we conduct interpretable Shapley additive explanations to analyze the model internal mechanism. The proposed IBTEM is compared with other popular machine learning methods and the experimental results show the IBTEM achieves outstanding superiority. Property value, fire spread and number of stories with minor damage are verified the most effective variables for loss prediction. In conclusion, the IBTEM realizes accurate and interpretable loss prediction of building fires, and assists relevant departments in making disaster relief decisions in a timelier manner.