Dynamic Spatiotemporal Straight-Flow Network for Efficient Learning and Accurate Forecasting in Traffic

To achieve accurate traffic forecasting, previous research has employed inner and outer aggregation for information aggregation, and attention mechanisms for heterogeneous spatiotemporal dependency learning, which results in inefficient model learning. While learning efficiency is critical due to the need for updating frequently the model to alleviate the impact of concept drift, limited work has focused on improving it. For efficient learning and accurate forecasting, this study proposes the dynamic spatiotemporal straight-flow network (DSTSFN). Breaking the aggregation paradigms employing both inner and outer aggregation, which may be redundant, the DSTSFN designs a straight-flow network that employs bipartite graphs to learn directly the dependencies between the source and target nodes for outer aggregation only. Instead of the attention mechanisms, the dynamic graphs/networks, which outdo static ones by possessing time-varying dependencies, are designed in the DSTSFN to distinguish the dependency heterogeneity, making the model relatively streamlined. Additionally, two learning strategies based on respectively the curriculum and transfer learning are developed to further improve the learning efficiency of the DSTSFN. Our study could be the first work designing the learning strategies for the multi-step traffic predictor based on dynamic spatiotemporal graphs. The learning efficiency and forecasting accuracy are demonstrated by experiments, which show that the DSTSFN can outperform not only the state-of-the-art (SOTA) predictor for accuracy by achieving a 2.27% improvement in accuracy and requiring only 8.98% of the average training time, but also the SOTA predictor for efficiency by achieving a 9.26% improvement in accuracy and requiring 91.68% of the average training time.