A data-driven supervised machine learning study for road-accident prediction in smart transportation operations and planning in smart cities

Road traffic accidents are still a key public-safety issue that drives the need for intelligent transportation systems to develop data-driven approaches to identifying passengers who are at risk of being involved in an accident. The aim of the current study is to compare machine-learning models on the classification of whether an accident occurs or not from the Transportation and Accident Prediction Dataset (TAPD). There are 5,500 cases and 17 attributes in the data set. The predictors listed were selected without leakage of the target, namely, so that there was no leakage into the target from the following variables: Accident_Severity and Alert_Generated. Accident_Occurred was selected as the binary target and Record_ID was excluded as an identifier. The classifiers that were evaluated and compared were KNearestNeighbors, Adaboost, CN2RuleInducer, SupportVectorMachine, RandomForest and NaiveBayes, with 10-fold cross validation. The accuracy, 95 % Wilson confidence intervals, precision, recall, specificity, F1 and balanced accuracy were used to measure performance. The highest accuracy (97.62%) and specificity (100%) was achieved by RF and NB, with both having a positive-class recall of 77.22%. The highest recall for the positive class was obtained by AdaBoost among the other ones (78.09%). The results show that the overall high accuracy can mask cases of accidents that are not captured, making it important to consider evaluation for class sensitivity in imbalanced safety data sets. TAPD is synthetic and the results are a controlled benchmark and not evidence of deployment readiness, therefore, validation with independent real-world traffic data is suggested.