Similarity Scoring with Random Field Models for Traffic Flow Management Applications

摘要

Despite the large archive of Traffic Flow Management (TFM) data, deriving non-trivial insights from this data via automation (e.g., machine learning, data mining) is notoriously hard. The difficulty is due, in part, to the high dimensionality of TFM data, the presence of both continuous and categorical variables, inconsistency in historical TFM strategies under apparently similar weather and traffic conditions, the dynamic nature of TFM, and the sheer quantity of historical TFM records. In this paper we propose a probabilistic Random Field model that addresses these (and other) issues by leveraging end-user inputs to supervise the learning of a similarity score for the robust comparison of TFM days. We demonstrate the utility and extensibility of the Random Field model and the similarity score it induces through preliminary clustering analyses.