Improving the prediction accuracy of the vehicle driver injury severity of accidents by topic modeling

One of the main challenges in reducing the number of road accidents is finding the causes of the accident. In the year, an average of 800,000 crashes occurs in Iran, which is ranked as the most deadly road accident in the world. This paper examines the accuracy of different models by using data mining methods due to the importance of identifying affecting factors on the injury severity of accidents. In this paper, we apply a topic modeling technique on some real-world database features to improve the accuracy of the vehicle driver injury severity prediction of accidents. In a double-layer data-driven framework, we first apply Latent Dirichlet Allocation (LDA), a generative probabilistic model to narrative features then we use simple, yet powerful classification algorithm called K-Nearest-Neighbors (KNN) and Decision Tree algorithm on the common U.S Federal Railway Administration (FRA) accident database for the period of 2010 - 2016 to identify vehicle driver injury severity factors of Highway-Railway Grade Crossing (HRGC) accidents. Experimental results show that by using topic modeling on narrative features, both algorithms have high accuracy in predicting the severity of the injuries.

Improving the prediction accuracy of the vehicle driver injury severity of accidents by topic modeling Keywords:

topic modeling , data mining , classification models , nearest neighbors , decision tree.