Duration prediction of urban freeway traffic accidents based on the M5P tree and hazard-based duration model

Among previous studies on accident duration prediction, the M5P algorithm has been shown to be an effective tool for predicting the clearance time. A limitation of the M5P, however, is that while the constructed tree-based models have multiple linear regression models as their leaves, the assumption of data normalcy of linear regression may not be suitable for traffic accident duration data, which is almost certainly nonsymmetrical. A hazardbased duration model (HBDM), on the other hand, is a good choice for time-to-event modeling situations. Unfortunately however, HBDMs are often used without prior clustering of the dataset, which aims at minimizing the data heterogeneity. As a result, useful insights into the data are often hidden. To address this, the current paper proposes a M5P-HBDM model for accident duration prediction, where the leaves of the M5P algorithm are replaced with HBDMs instead of linear regression models. The new model is then tested on a freeway accident dataset including 602 accident records along a segment of I-64 in Virginia, over the period from 01/01/2005 to 12/31/2006. The first 500 records are used to train the M5P tree, HBDM, and the M5P-HBDM, and the remainders (i.e., 102 records) are used to compare the performance of the three different kinds of accident duration prediction models. The results show that M5P-HBDM performs much better than both M5P and HBDM.