Adaptive Gaussian mixture model for identifying outliers in historical route travel times

Abstract

AbstractThe historical route travel time data contain a few outliers composed of subpopulations signifying unusual traffic conditions. These outliers compose the right tail of the distribution against the central part occupied by most normal travel times dominating usual traffic conditions likely to be composed of subpopulations. The diverse subpopulations of two types of travel times result in the structural random changes of distribution shapes. This creates the problem for fixing the boundary between the two types of travel times. To address the problem, an adaptive Gaussian mixture model was formulated with two types of components, and an algorithm was put forward to determine the critical number of components in an iterative manner by adapting two types of components to provide an adequate fit to the two parts of the distribution respectively. The determined two types of components can not only fix the boundary to identify outliers reasonably, but also quantify the latent subpopulations of two types of route travel times. Thus, the route‐level travel time variability can be measured under usual/unusual traffic conditions. Two kinds of data were used to illustrate the good effects on the identification of outliers and to demonstrate the vital role of outliers in the measure of variation of route travel time.