Robust Prediction of Traffic Flow Based on Multi-Task Graph Convolution Network

Abstract

Accurate traffic prediction is indispensable to realize an intelligent transportation system in a smart city. However, the traffic network has a complex spatial structure, and short-term / long-term time dependence. Besides, the traffic network is faced with the various external factors, including the instability of hardware equipment in data collection and network time delay in the process of data transmission. And most of the existing methods predict the traffic flow under the traffic data is complete, which lose the tolerance in the case of abnormal data. In this paper, to improve the robustness of the traffic flow prediction model, we propose a multi-task graph convolution network (MTGCN) structure, which combines multi-task learning (MTL) with graph convolution network (GCN). Firstly, the GCN method is used to obtain the spatiotemporal correlation of traffic network. Secondly, we use a model that is able to do MTL via multiple outputs, each corresponding to the same traffic network at different adjacent time durations. Experiments on real traffic flow data sets demonstrate the effectiveness and robustness of the proposed method in the presence of abnormal patterns such as Gaussian noise and random missing.