Outlier Analysis to Improve the Performance of an Incident Duration Estimation and Incident Management System

Traffic incidents, as non-recurrent events, are one of the major causes of congestion in the transportation network. To mitigate the impacts of such incidents and to recover the performance of transportation systems as safely and quickly as possible, most responsible agencies over the past decades have implemented various traffic incident management systems, and an incident duration prediction model is one of the key components to estimate the impact of time-varying incidents on the network. Many studies have been undertaken to develop a robust prediction model of incident duration, but they have struggled to provide a reliable and accurate estimation result because of various data and modeling issues, such as a highly skewed distribution, complex correlations, heteroscedasticity, many outliers, and so forth. This study proposes an outlier analysis process for estimating the outlier-ness of each detected incident and utilizing such outlier information to improve accuracy of prediction of incident duration estimation and detect any system deficiency. An ensemble modeling technique and various outlier detection methodologies have been used to estimate the outlier-ness, and a hybrid association rule mining method has been applied to classify the detected outliers as anomalies or noises. Lastly, through the model evaluation and application example in this study, we can conclude that the proposed outlier analysis process can improve the accuracy of incident duration estimation and detect the potential system deficiencies associated with incident response, data recording, resource management, and so forth.