Graph Convolutional Network with Gated Multi-mode Fusion for Traffic Forecasting

As a crucial problem in Intelligent Transportation Systems (ITS), traffic forecasting has attracted attention from an increasing number of researchers in recent years. Currently, the most promising strategy is spatio-temporal graph neural networks, which leverage graph neural networks for spatial dependency and sequence learning modules for temporal dynamics simultaneously. However, previous studies omit that there are complex relations between multi-mode features. And the temporal patterns captured by previous models are always limited or time-consuming. To address the above issues, this paper proposes Multi-mode Fusion Graph Neural Network (MFGNN). Different from the majority of previous studies that simply concatenate different features together, the proposed method adopts a framework that explicitly considers relations between primary features and auxiliary features. To handle the complex relations between different features, a gate-based fusion module is specially designed to filter unnecessary information. The module can control how much information from the auxiliary branch is fused and learned automatically with the data. Moreover, a graph-based temporal relation learning module is proposed. Different from convolutional neural networks or recurrent neural networks, the module establishes parameterized relations between every pair of input time slots directly, which can learn flexible yet efficient temporal patterns of traffic. Extensive experiments are conducted on three real-world datasets and two popular traffic forecasting tasks. The experimental results demonstrate the superiority of the proposed method.