Extralonger: Toward a Unified Perspective of Spatial-Temporal Factors for Extra-Long-Term Traffic Forecasting

TL;DR

Drawing inspiration from Albert Einstein's relativity theory, which suggests space and time are unified and inseparable, Extralonger is introduced, which unifies temporal and spatial factors and sets new standards in long-term and extra-long-term scenarios.

Abstract

Traffic forecasting plays a key role in Intelligent Transportation Systems, and significant strides have been made in this field. However, most existing methods can only predict up to four hours in the future, which doesn't quite meet real-world demands. we identify that the prediction horizon is limited to a few hours mainly due to the separation of temporal and spatial factors, which results in high complexity. Drawing inspiration from Albert Einstein's relativity theory, which suggests space and time are unified and inseparable, we introduce Extralonger, which unifies temporal and spatial factors. Extralonger notably extends the prediction horizon to a week on real-world benchmarks, demonstrating superior efficiency in the training time, inference time, and memory usage. It sets new standards in long-term and extra-long-term scenarios. The code is available at https://github.com/PlanckChang/Extralonger.

Figures (8)

• Figure 1: Overview of the architecture.
• Figure 2: Classical pipeline (TOP) and New pipeline (BOTTOM).
• Figure 3: Given the yellow circle and star are the analysis target at $t$ time step, (a) only Extralonger aggregates along different nodes and different time steps globally; (b) CNN-based methods aggregate the information within the restricted receptive field (red box), and the other early methods aggregate along the temporal and spatial dimensions separately (two orange belts). Capturing temporal and spatial dependency globally with the Unified Spatial-Temporal Representation (whole area), Extralonger does simultaneously and efficiently.
• Figure 4: Global-Local Spatial Transformer.
• Figure 5: Resource cost comparison w.r.t. memory usage (LEFT), training time (MIDDLE) and inference time (RIGHT). The y-axis represents the logarithm of the value. The dashed line reveals the fitted trendline based on actual results.