Building sparse models for traffic flow prediction: an empirical comparison between statistical heuristics and geometric heuristics for Bayesian network approaches

Abstract

As the sensors become pervasive in transportation systems, traffic prediction becomes more challenging due to the significant increase of data. In this paper, we study how to design BNs for traffic flow prediction with an emphasis on node selection strategy. Three node selection strategies are evaluated based on PeMS (Performance Measurement System) dataset: the Graphical Lasso approach, geographically neighboring approach, and direct correlation approach. Test results show that feeding all the available data into the prediction model might not be a wise choice. The three node selection strategies mentioned above are generally equivalent in prediction performance. We also find that direct correlation approach may lead a sparser model and meanwhile keeps good prediction performance on the PeMS dataset. The obtained conclusions not only help simplify the design of traffic flow prediction models in practice, but also shed light on how to effectively use the massive data.