A Hybrid Machine Learning Approach for Performing Predictive Analytics on Road Accidents

Abstract

Nineteen thousand accidents occurred on the roads of the Indian state of Andhra Pradesh in the year 2020. Every year, traffic inconsistencies are a major cause of fatalities, injuries, and property damage. Road and traffic accidents are uncertain, unpredictable events, and understanding the elements that influence them is necessary for their prevention. Age, gender, location, time, climate, and many more factors are among those linked to traffic accidents. Several variables, most of which are discrete, define road accidents. The proposed framework applies statistical modeling and machine learning to the accident dataset containing details of 18,251 road accidents that took place in the Vijayawada region over the period 2013 to 2022, to find the hidden patterns particular to the severity of accidents. These patterns will exhibit common features of accidents that took place at a black spot. The obtained insights would help reduce further accidents. The varied nature of accident data analysis is the main challenge. As a result, heterogeneity must be considered while analyzing the data, or any relationship between the data may be obscured. The whole accident data is divided into zones and the proposed methodology is applied to each zone individually. To overcome the heterogeneity of data, gaussian mixture modeling is used to perform soft clustering. Then making use of results, like the means and variances obtained from each gaussian, a training dataset is created, and accidents are classified based on the severity of consequence using a support vector classifier, decision tree classifier, random forest, K-Nearst Neighbours classifier, and naive bayes classifier. Later accuracies of all algorithms are compared to find the best suitable model for a zone. The proposed framework makes use of python language libraries like pandas, sklearn, matplotlib, and joblib.