Can we Learn from Outliers? Unsupervised Optimization of Intelligent Vehicle Traffic Management Systems

Vehicle traffic flow prediction is an essential task for several applications including city planning, traffic congestion management and smart traffic light control systems. However, recent solutions suffer in outlier situations where traffic flow becomes more challenging to predict. In this work, we address the problem of predicting traffic flow on different intersections in a traffic network under the realistic assumption of having outliers. Our framework, called OBIS, applies an existing LOF-based approach to detect outliers on each intersection in the network separately. Based on the spatio-temporal interdependencies of these outliers, we infer the correlations between intersections in the network. We use these outlier-based correlations then to improve the predictability of existing traffic flow prediction systems by selecting more relevant inputs for the prediction system. We show that our framework considerably improves the performance of LSTM-based models both under outlier scenarios and also under normal traffic. We test our framework under two real-life settings. In the first, we show how improving the predictability using our framework reduces the overall delays of vehicles on an intersection with a smart traffic light control system. In the second, we demonstrate how OBIS improves the predictability of a real dataset from four trajectories of intersections in the city of The Hague. We share the latter dataset together with an implementation of our framework.