Using Graph Attention Network to Predicte Urban Traffic Flow

Abstract

In recent years, deep learning has been widely used in image classification and natural language processing and has achieved quite good results. Graphs are a very important part of computer algorithms and data structures. Graph structures can find lost of mappings in natural world. It emphasizes the relationship between nodes and link, and this property is widely used in social network processing, risk control, cyber security, and smart cities. The traffic flow problem is an important factor that plagues urban development. Large cities need to consider many factors when planning new construction, such as the distribution of residential communities, natural topography, and even underground pipelines. The city construction process adjusts the current traffic flow, inevitably impacting the original urban transportation network. The uncertainty of such impact is one of the reasons why many urban construction schemes are difficult to advance. Here we propose a graph attention mechanism-based urban traffic flow prediction, and we propose a method that separates the urban grid layout and traffic calculation to handle remapping. Our model can accurately predict both global and local traffic in cities while achieving satisfactory results in regression evaluation metrics, and our model is also a lightweight model that provides a basis for future research on small-scale devices.