Fine-Grained Spatial-Temporal Representation Learning with Missing Data Completion for Traffic Flow Prediction

Spatial-temporal traffic flow prediction is beneficial for controlling traffic and saving traffic time. Researchers have proposed prediction models based on spatial-temporal representation learning. Although these models have achieved better performance than traditional methods, they seldom consider several essential aspects: 1) distances and directions from the spatial aspect, 2) the bi-relation among historical time intervals from the temporal aspect, and 3) missing historical traffic data, which leads to an imprecise spatial-temporal features extraction. To this end, we propose Fine-Grained Features learning based on Transformer-encoder and Graph convolutional networks (FGFTG) to improve the performance of traffic flow prediction in a missing data scenario. FGFTG consists of two components: feature extractors and a data completer. The feature extractors learn fine-grained spatial-temporal representations from spatial and temporal perspectives. They extract smoother representation with the information of distance and direction from a spatial perspective based on graph convolutional networks and node2vec and achieve bidirectional learning for temporal perspective utilizing transformer encoder. The data completer simulates the traffic flow data distribution and generates reliable data to fill in missing data based on generative adversarial networks. Experiments on two public datasets demonstrate the effectiveness of our approach over the state-of-the-art methods.