Accident severity prediction in big data using auto-machine learning

Abstract

Estimating the severity of a traffic accident is a problem in motor vehicle traffic because it affects saving human life. If the severity value can be predicted before the accident, all the emergency teams needed could be sent to the area to provide faster first aid. With this aim, we studied a big data set for accidents in the United States of America (USA) between 2016 and 2020, almost 2.25×106 rows long. First, the data is preprocessed by removing unnecessary variables. Then rows with blank cells are removed. Finally, about 1.7×106 rows length data are left for the prediction process. A Machine Learning (ML) algorithm has been used to classify the severity based on 16 input parameters. Binary-to-decimal count conversation has been used as a novel preprocessing method. As a result, the model has been built with a total accuracy of 0.816. The test results are also validated with precision, recall, and f1-score values. An Auto-Machine Learning (Auto-ML) model has been developed and trained to predict the severity of a possible traffic accident based on the weather and road conditions. Thus, it will be possible to direct emergency units to areas with high accident severity, and preventing a fatality.