Spatio-temporal graph neural networks for missing data completion in traffic prediction
Jiahui Chen, Lina Yang, Yi Yang, Ling Peng, Xingtong Ge
Abstract
Missing traffic data completion is a key part of the construction of a smart city. However, due to cost constraints and other reasons, many locations do not have sensors to record traffic data. Most research methods do not systematically consider filling in missing traffic data. This study explores a new spatio-temporal feature extraction layer that includes spatio-temporal feature fusion, graph learning on an adaptive adjacency matrix, and a gated recurrent unit with a mask for missing traffic data completion. This idea is based on a hypothesis: missing data can be inferred from the spatio-temporal features of other nearby recorded sensor nodes. Therefore, we propose an end-to-end traffic model dealing with missing data - missing traffic data completion graph neural networks (MTC-GNN). Experiments demonstrate that the proposed model can learn spatio-temporal patterns and fill in speed from traffic data with various missing ratios and outperform existing models.