Validation of Machine Learning Algorithms as Predictive Tool in the Road Safety Management Process: Case of Network Screening

Abstract

Safety performance functions (SPFs) are the key regression tools in the road safety management process (RSMP) and are used to predict crash frequency given a set of roadway and traffic factors. Although regression-based SPFs have been proven to be reliable tools for road safety predictive analytics, some limitations and constrains have been highlighted in the literature, such as the need to assume a probability distribution, the need to select a predefined functional form, possible correlation between independent variables, and possible transferability issues. An alternative to traditional regression models as predictive tools is the use of machine learning (ML) algorithms. This research compared the prediction performance of three well-known ML algorithms, i.e., support vector machine (SVM), decision tree (DT), and random forest (RF), with that of traditional SPFs, and applied and validated ML algorithms in network screening, which is the first step in the RSMP. To achieve these objectives, traditional SPFs using negative-binomial (NB) generalized linear regression were estimated and compared with ML algorithms using three different goodness-of -fit criteria. A data set of urban signalized and unsignalized intersections from two major municipalities in Saskatchewan (Canada) was considered as a case study. Ranking consistency tests of collision-prone locations identified using ML-based and SPF-based performance measures were conducted. The results showed that the consistency of ML-based measures in identifying hotspots was comparable to that of SPF-based measures, particularly the excess (predicted and expected) average crash frequency. Overall, the results of this research support the use of SVM, DT, and RF as predictive tools in network screening.