Evaluating the Performance of Algorithms for the Detection of Travel Time Outliers

Several technologies—including automatic license plate readers, cell phone probes, dedicated Global Positioning System probes, the automatic identification of vehicles equipped with transponders or toll tags, and Bluetooth detectors—can acquire vehicle travel times. The travel times measured by all these technologies contain errors and biases. As a result, several filtering and outlier detection algorithms have been proposed to identify erroneous data and to exclude them from the analysis. However, it is difficult to assess the performance of an outlier detection algorithm in absolute terms or relative to any other algorithm through the use of field data because the true travel times are unknown. Furthermore, it is not possible to identify how well the algorithm can identify a given source of outliers. In this paper a framework is proposed for the evaluation of algorithms that detect travel time outliers. The framework can be customized to address the specific characteristics of any travel time sensor technology. However, in this paper the framework is demonstrated through its application to travel times acquired by Bluetooth detectors on arterial roadways. The framework is used to evaluate the performance characteristics of two outlier detection algorithms proposed by Dion and Rakha. The results show the performance characteristics of both algorithms for a wide range of operating conditions. One of the algorithms is shown to have an approximately 30% likelihood of providing worse results than if no outlier detection algorithm is used. The other algorithm is shown to provide improvements under almost all conditions, with a relative improvement of up to 60%.