Unlocking the Full Potential of Deep Learning in Traffic Forecasting Through Road Network Representations: A Critical Review

Abstract

AbstractResearch in short-term traffic forecasting has been blooming in recent years due to its significant implications in traffic management and intelligent transportation systems. The unprecedented advancements in deep learning have provided immense opportunities to leverage traffic data sensed from various locations of the road network, yet significantly increased the models’ complexity and data and computational requirements, limiting the actionability of the models. Consequently, the meaningful representation of traffic flow data and the road network has been highlighted as a key challenge in improving the efficiency, as well as the accuracy and reliability of forecasting models. This paper provides a systematic review of literature dedicated to spatiotemporal traffic forecasting. Three main representation approaches are identified, namely the stacked vector, image/grid, and graph, and are critically analyzed and compared in relation to their efficiency, accuracy and associated modeling techniques. Based on the findings, future research directions in traffic forecasting are proposed, aiming to increase the adoption of the developed models in real-world applications.