TSTNet: A Sequence to Sequence Transformer Network for Spatial-Temporal Traffic Prediction

Making accurate traffic forecasting is of great importance in smart city-related researches. However, as the traffic features like traffic speed have a complex spatial-temporal characteristics, how to build an accurate traffic prediction model is still an open challenge. In this work, we propose TSTNet, a Sequence to Sequence (Seq2Seq) spatial-temporal traffic prediction model. TSTNet adopts Graph Attention Network (GAT), which can learn the spatial feature aggregation, to build spatial dependency. For temporal dependency, TSTNet applies a Seq2Seq Transformer structure to establish temporal dependency. As a GAT layer’s operation only aggregate the attribute information for neighbor nodes, it does not involve any spatial positional information. Similarly, if we apply the Transformer model on sequence learning tasks, the Transformer model also does not involve any temporal positional information as it does not know the exact time slot of different inputs. To solve the above problems, TSTNet implements a spatial-temporal embedding method to obtain the spatial-temporal positional representation for each input data. We evaluate TSTNet on traffic speed prediction tasks with other baselines upon two real-world datasets, the results show that TSTNet outperforms all the baseline models