A Causal Deep Learning Framework for Traffic Forecasting

Inferring causal relationships from data has the potential to significantly enhance traffic forecasting and management. However, causality is often neglected in recent literature, due to the demanding processes required to infer causal links between traffic variables. In this work we resort to the novel Neural Granger method to detect the causality structure of the road network traffic of the Athens city center (Greece) based on data monitored by loop detectors. Furthermore, we show the impact of the detected causalities on the forecasting performance of hourly volumes of traffic flow data. The detected causal relations reveal the existence of strong daily traffic patterns and dependencies between locations at the perimeter and in the center of the city. In addition, the detected causal relationships allow for more efficient and accurate forecasting of future traffic conditions.