01626316

Component

The crash likelihood prediction has been an important issue in traffic safety studies. The problem of missing data or imperfect information may have a negative impact on the accuracy of crash likelihood prediction models. In this paper, a novel approach is proposed that combines the probabilistic principal component analysis (PPCA) for missing data imputation and support vector machines (SVMs) as the crash likelihood prediction model. To avoid the potential overfitting issue, the backward sequential feature selection is conducted to select the optimal combination of explanatory variables. To verify how the PPCA method affects the prediction accuracy of SVMs with 3 kinds of kernels, i.e., linear, Gaussian, and polynomial, the proposed approach is applied to a field study on an urban expressway. The models are trained and tested with 123 crash records and 5-month traffic flow data. The 5-fold cross validation is employed to train the classifier and verify its prediction accuracy under different percentages of missing information. Numerical results show that SVM models with full explanatory variables without feature selection result in similar values as the missing ratio increases from 0 to 40%, in terms of the area under the curve (AUC) of receiving operation characteristic (ROC). It verifies the stable performance of PPCA in missing data imputation for the crash likelihood prediction. On the other hand, SVM models established on the basis of optimal variables, selected by the backward sequential feature selection, show unsatisfactory AUC values with the increase of the missing rate although they provide better prediction performance than SVM models with full variables when the missing ratio is close to zero. It indicates that the selection of explanatory variables and PPCA-based missing data imputation may not be implemented simultaneously when the missing rate of observations is high. The approach of incorporating missing information imputation with crash likelihood prediction is generic and can be applied to other machine learning classification models.

This paper was sponsored by TRB committee ANB20 Standing Committee on Safety Data, Analysis and Evaluation.

01618707

17-04420

English

21p

2017

Transportation Research Board 96th Annual Meeting

Digital/other

Appendices; Figures; Maps; References (25) ; Tables

Crash risk forecasting; Data quality; Mathematical prediction; Traffic safety

Principal Component Analysis

Support vector machines

Highways; Safety and Human Factors

Transportation Research Board Annual Meeting 2017 Paper #17-04420

TRIS, TRB, ATRI

Dec 8 2016 11:42AM