An accident prediction approach based on XGBoost

Abstract

As an important threat to public security, urban fire accident causes huge economic loss and catastrophic collapse. Predicting and analyzing the interior rule of urban fire accident from its appearance needed to be solved in the field. In this paper, we propose a new urban fire accident prediction approach based on XGBoost. The method determines the predictive indexes in a quantitative and qualitative way from different characteristics in various kinds of fire accidents. For screening the features we need, we adopt the feature selection algorithm based on association rules. For data cleaning, we use a method based on Box-Cox transformation that transforms the continual response variables from the feature space for removing the dependencies on unobservable errors and the predictor variable to some extent. Then we use the data to train the model based on XGBoost to obtain the best prediction accuracy. Experiments show that the method provides a feasible solution to urban fire accident prediction. The method contributes to improving the public security situation, we have added the method and related model to the City in a box™, Shenzhen Aerospace Smart City System Technology Co., Ltd.