MFAGCN: Multi-Feature Based Attention Graph Convolutional Network for Traffic Prediction

Large-scale traffic data mining provides a new solution to alleviate traffic congestion and improves traffic service. As an important part of traffic data analysis, traffic multi-feature prediction is widely concerned, and several machine learning algorithms are also applied in this field. However, this is very challenging, because each traffic feature has a highly nonlinear and complex pattern, and there are great differences between multiple features. A large number of the existing traffic feature prediction methods focus on extracting a single traffic feature and lack the ability of analyzing multiple features. This paper proposes a new multi-feature based attention graph convolutional network (MFAGCN) to solve the problem of the prediction of multiple features in traffic, the proposed method has 4.53% improved to the conventional methods. The three features predicted by MFAGCN are traffic flow, occupancy rate, and vehicle speed. Each feature is modeled with three temporal attributes, namely weekly period, daily period, and nearest period. In this paper, multi-feature prediction mainly includes two parts: (1) Establish a spatiotemporal attention mechanism for capturing the dynamic spatiotemporal correlation of multiple features. (2) Different convolutional kernels and activation functions are used for each feature after splitting. The experimental results on real traffic datasets have verified the effectiveness of the MFAGCN .