Abstract: With the rapid changes in city traffic safety, there is a need to update macro safety models to predict crashes accurately at various times. Two main challenges: the homogeneous datasets and effective data collection for timely updating, have hindered researchers' ability to update the models, however. This study applied boosting techniques, which are well adapted to the conditions of data heterogeneity and small sample size, to macro safety model updating. To this end, crashes and regional characteristics were collected in 2009 and 2016 for Shanghai, China, as the source and target data domains, respectively. Four boosting-based updating models, AdaBoost.R2, two-stage TrAdaBoost.R2, Gradient Boosting, and CatBoost (an abbreviation for categorical boosting), along with a traditional two-stage Bayesian updating model, were established to evaluate and compare crash-prediction performance by Root Mean Square Error. The results showed that the CatBoost algorithm, with its ability to cope with heterogeneous datasets and categorical features, outperformed all the other methods. A further investigation into the optimal target sample size analysis was conducted. The three advanced boosting algorithms tended to have similar results around the proportion of 40% of target data (105 TAZs) in the training dataset. The two-stage TrAdaBoost.R2 and CatBoost tended to outperform in the near-full sample size and small target sample size, respectively. Thus, the CatBoost algorithm model with 40% target data is recommended for macro safety model updating. These findings can be applied to the practice of long-term timely traffic safety monitoring and data collection optimization.
Zhicheng Dai, Xuesong Wang*, Xiaohan Yang. Macro-level Safety Model Updating: Application of Boosting Techniques. Transportation Research Board 101st Annual Meeting, Washington D.C., USA, 2022. 1.9-13.