Data-driven kitchen fire prediction based on environmental variables

Efficient prediction is crucial to preventing harm caused by kitchen fires. In this paper, we propose a kitchen fire model using the data collected by the Twente Fire Brigade. Specifically, we utilize the permutation techniques of random forests and perform classic stepwise regression methods to select the explainable environmental variables. For unstable results, we propose stabilization methods. Moreover, we build a Poisson generalized linear model which successfully captures the spatial patterns seen in the data.