Comparative analysis of machine learning techniques for structural fire prediction models

Abstract

Structural fires frequently occur, causing injuries, loss of property, and taking lives. The environment where fires can easily occur can be seen through historical data, and it is desirable to prevent them in advance through predictive models. This study builds various machine learning models to predict the fire and performs a comparative analysis of the models. We compare decision trees, random forest, extra tree classifier, XGBoost, neural networks, Stochastic Gradient Descent Classifier (SGDC), Gradient Boosting Method (GBM), Light Gradient Boosting Method (LGBM), and adaboosting. We use the data obtained from the local fire department in South Korea to build a fire prediction model. Before creating a fire prediction model, we analyze and extract the factors affecting the fire to perform machine learning. We implement a comparative analysis and show accuracy, precision, recall, and F1-score for the fire prediction models. The predictive model includes the models that are suited for fire prediction. We show the computational times and F1-Scores of each prediction model. The prediction model built with the random forest is the most accurate and precise, and there is a little difference in the accuracy of each model trained with the extra tree classifier, XGboost, and neural network. For the F1 Score, the model with a neural network shows the best value. Key attributes affecting the fire are identified. Since prediction models are built and performed comparative analysis on only data from specific regions, it is difficult to apply in general in all areas. © 2020 Alpha Publishers. All rights reserved.