Transfer Knowledge Between Sub-Regions For Traffic Prediction Using Deep Learning Method

In modern society, traffic is a very important aspect of our social lives, but a lot of problems arise such as traffic accident, congestion, air pollution, etc. The emergence of ITS (Intelligent Transportation System) brings a range of scientific ways to tackle these problems, among which there is a crucial need to study prediction methods in modern traffic systems. Fortunately, big data in ITS and advanced machine learning technologies motivate us to implement extensive researches in this area. In this paper, we introduce an innovative method for traffic prediction. We split the whole traffic system into separate sub-regions. Then, each of them would be modeled by deep graph neural networks, which could extract unique characters of the region as well as preserve the topology property of it. Moreover, we transfer the useful knowledge between these regions and share the basic information in order to improve the performance of our model. Finally, we conduct experiment on real world data set and achieve the best performance, which proves the effectiveness of our method.