A proactive decision support system for predicting traffic crash events: A critical analysis of imbalanced class distribution

摘要

Real-time crash prediction plays a key role in enhancing traffic safety as well as mitigating disruptions to road users. The further improvements of predictability require the systemic analysis of crash likelihood within the driver–vehicle–environment triptych. This study presents a proactive decision support system that can predict crash events based on vehicle kinematics, driver inputs, roadway geometric features and real-time weather data. Modeling approaches that rely on Random forest, Support Vector Machine and Multilayer Perceptron machine learning techniques were applied to establish efficient crash predictions. Moreover, crash events are generally unexpected and occur rarely, thus classification results can yield deceivingly high prediction performance which are usually driven by the majority class at the expense of having poor performance on the crucial minority class. Therefore, this paper attempts to add to the current knowledge by investigating crash likelihood based on compared different data balancing techniques to improve the predictive performance through three balancing techniques: over-sampling, under-sampling and synthetic minority over-sampling (SMOTE). The highest performances have been acquired using SMOTE strategy as MLP achieved a 94.5% precision, 94.2% f1-score, 93.7% AUC and 95.3% recall, while SVM achieved a 91.5% g-mean. Furthermore, results indicated that more than 62% of total crashes have been reported in downhills and curved downhills, and 44% of all crash instances have been reported during both snow and rain weather patterns. Overall, the findings highlighted the significance of the explanatory variables associated with potential crash events and can suggest to decision-makers a safe and credible system for enhancing traffic safety.