Coupled Layer-wise Graph Convolution for Transportation Demand Prediction

• DOI: 10.1609/aaai.v35i5.16591
• Publication Date: 2022-09-08T18:27:57Z
• Authors: Junchen Ye; Leilei Sun; Bowen Du; Yanjie Fu; Hui Xiong
• Keywords: Adjacency matrix; Adjacency list; Convolution (computer science)

Graph Convolutional Network (GCN) has been widely applied in transportation demand prediction due to its excellent ability capture non-Euclidean spatial dependence among station-level or regional demands. However, most of the existing research, graph convolution was implemented on a heuristically generated adjacency matrix, which could neither reflect real relationships stations accurately, nor multi-level demands adaptively. To cope with above problems, this paper provides novel convolutional network for prediction. Firstly, architecture is proposed, different matrices layers and all are self-learned during training process. Secondly, layer-wise coupling mechanism provided, associates upper-level matrix lower-level one. It also reduces scale parameters our model. Lastly, unitary constructed give final result by integrating hidden states gated recurrent unit, temporal dynamics simultaneously. Experiments have conducted two real-world datasets, NYC Citi Bike Taxi, results demonstrate superiority model over state-of-the-art ones.