Identification of Influencing Attributes for the Traffic Accident Severity Prediction Model

Abstract

This paper deals with identification of the most influencing input attributes related to the accuracy of the prediction model. It is assumed that the prediction model may be represented by any machine learning-based models, including artificial neural networks, fuzzy models, etc. Selection of influencing attributes is based on particle swarm optimization (PSO) combined with neural networks. The role of neural networks is to estimate fitness of each particle during the search procedure implemented using a PSO algorithm. The presented feature selection method represents the first step in the prediction model design which is applied on the accident severity prediction. The method is applied on a dataset characterized with weak correlation between the model’s inputs and output representing an accident severity. The proposed approach has shown the ability to identify subsets of factors that have an influence on accident severity prediction. Selection of appropriate inputs improves the prediction model accuracy.