Traffic Speed Data Investigation with Hierarchical Modeling

Abstract

This paper presents a novel topic model for traffic speed analysis in the urban environment. Our topic model is special in that the parameters for encoding the following two domain-specific aspects of traffic speeds are introduced. First, traffic speeds are measured by the sensors each having a fixed location. Therefore, it is likely that similar measurements will be given by the sensors located close to each other. Second, traffic speeds show a 24-hour periodicity. Therefore, it is likely that similar measurements will be given at the same time point on different days. We model these two aspects with Gaussian process priors and make topic probabilities location- and time-dependent. In this manner, our model utilizes the metadata of the traffic speed data. We offer a slice sampling to achieve less approximation than variational Bayesian inferences. We present an experimental result where we use the traffic speed data provided by New York City.