Beyond Short-term Traffic Forecasting Models: Navigating Through Data Availability Constraints

Abstract

Short-term traffic forecasting supports route planning and decision making before traffic congestion occurs. Thanks to its direct application in real-world scenarios, short-term traffic forecasting remains as one of the hot topics within research on Intelligent Transportation Systems. During the last decade, researchers have heavily focused on proposing advanced and complex modeling techniques based on Deep Learning architectures. Motivated by the revolution Deep Learning has supposed to computer vision and natural language processing, authors continuously evaluate state-of-the-art methods on traffic forecasting datasets. However, published performance improvements are narrow. This Thesis conducts first a literature review on short-term traffic forecasting models, intending to shift the community research efforts beyond increasing the accuracy of traffic predictions. The experience accumulated in the course of the presented survey, allows drawing a road map of challenges and research opportunities for the years to come. Aiming to lead by example, several of the above challenges are directly addressed in this Thesis, in detail those that gravitate around different levels of data constraints.Scholars rely on extensive datasets for adjusting proposed models, however, reality differs from these ideal experimental setups. Three levels of data availability are analyzed: 1) traffic measurements collected during a whole year; 2) traffic surveillance limited to a few weeks; 3) no traffic data available for a particular location. The first experimental setup aims to demonstrate that increasing the complexity of the models used for short-term traffic forecasting does not yield more accurate predictions in those scenarios where traffic recordings are accessible. The second case study explores how to learn traffic forecasting models under limited data holdouts, while maintaining a similar predictive performance regarding those models built without any data constraint. The last and most challenging scenario, delves into the characterization of sensorless locations. This research path has the potential of reducing the number of sensors permanently installed across a traffic network, by pairing sensorless road segments to those roads that share a similar traffic behavior.