Development of Incident Impact Prediction Models: A Comparison using Bagging and Boosting Techniques

Abstract: Traffic incidents have long been recognized as a major congestion contributor and a significant threat to urban mobility as well as safety. In this regard, prediction of incident impact on the surrounding road network has become an important task which may dictate the deployment of appropriate traffic management measures in real time to reduce the impact up to an acceptable level. This study aims to bring insights into the development of traffic incident impact prediction models as functions of incident characteristics and traffic management strategies using multiple supervised machine learning techniques. First, traffic incident is characterized using several key variables and multiple performance measures (MOEs) representing incident impact are selected. Then various traffic management strategies are formulated based on diversion routing for different incident scenarios. These incident scenarios are then simulated using a calibrated and validated VISSIM micro-simulation model with reference to an arterial road network in Kolkata City, India. It is followed by an evaluation of the alternative traffic management strategies on the basis of the simulation results. Then several supervised machine learning (ML) models were developed to predict the incident impacts under different incident and traffic management scenarios using the evaluation results obtained from the VISSIM micro-simulation. In this study, we intend to develop impact prediction models using a ‘Bagging’ as well as a ‘Boosting’ ML techniques – ‘Random Forest (RF)’ and ‘Light Gradient Boosting Machine (LightGBM)’ are used as the bagging and boosting methods respectively. A series of features including incident characterizing variables, location characterizing variables, and variables representing traffic management strategies are employed as the predictor variables, while the set of MOEs representing impact of incident are used as the target variables. Separate models are developed for all the four MOEs considering both peak and off-peak hour conditions using the ML techniques. It is observed that both the ML techniques can give reasonable prediction results. The ‘Boosting’ models are observed to predict incident impact more accurately as compared to the ‘Bagging’ models. This method of developing incident impact prediction models may be used advantageously by researchers and practitioners for various application contexts. Though the VISSIM micro-simulation results provide an extensive evaluation analysis of the traffic management strategies under different incident scenarios, formulation of a generalized model may further enhance the scopes of the methodological approach. In case of an incident occurrence, the traffic authority may just use the prediction models to assess the effectiveness of alternative traffic management strategies and identify the most suitable strategy based on the predicted incident impact, without going for more rigorous micro-simulation analysis.