Predicting the Effect of an Accident on Traffc Congestion

Abstract

In this study a Multilayer Perceptron was applied to predict tra c jams based on accidents. Emergency service messages were used as an indication of an accident. The model uses emergency service messages and road tra c information as input features. A dataset was created that com- bines information about accidents with tra c data. The model predicts if an accident is followed by a tra c jam within the next 15 minutes. The results show that a higher AUC than the base- line of 0.5 is achievable. Furthermore two di erent techniques to deal with unbalanced data were explored, namely resampling and cost-sensitive learning.