Exploring the potential of data mining techniques for the analysis of accident patterns
Carlo Giacomo Prato, Shlomo Bekhor, Ayelet Gal‐Tzur, David Mahalel, Joseph N. Prashker · 2010
Research in road safety faces major challenges: individuation of the most significant determinants of traffic accidents, recognition of the most recurrent accident patterns, and allocation of resources necessary to address the most relevant issues. This paper intends to comprehend which data mining techniques appear more suitable for the objective of providing a broad picture of the road safety situation and individuating specific problems that the allocation of resources should address first. Descriptive (i.e., K-means and Kohonen clustering) and predictive (i.e., decision trees, neural networks and association rules) data mining techniques are implemented for the analysis of traffic accidents occurred in Israel between 2001 and 2004. Results show that descriptive techniques are useful to classify the large amount of analyzed accidents, even though introduce problems with respect to the clear-cut definition of the clusters and the triviality of the description of the main accident characteristics. Results also show that prediction techniques present problems with respect to the large number of rules produced by decision trees, the interpretation of neural network results in terms of relative importance of input and intermediate neurons, and the relative importance of hundreds of association rules. Further research should investigate whether limiting the analysis to fatal accidents would simplify the task of data mining techniques in recognizing accident patterns without the “noise” probably created by considering also severe and light injury accidents.