Memory-facilitated joint-space shift adaptation in traffic forecasting

Traffic forecasting, crucial for intelligent transport systems, faces significant challenges from distribution shifts due to the dynamic nature of traffic patterns. Although normalisation approaches have been proposed to address distribution shifts in other time-series forecasting tasks, e.g., predicting electricity consumption load and estimating influenza-like illness patient numbers, they fall short in handling the complex spatial and temporal shifts in traffic data. In this paper, we propose a novel memory-facilitated joint-space shift adaptation framework, namely ST-Align, to address the problem in traffic forecasting.

ST-Align comprises two key components targeting the input and latent space, respectively: a memory-based data alignment module in the input-space, and an end-to-end memory network structure dedicated to alignment within the latent-space. This joint-space design enables our ST-Align framework to effectively capture and adapt to dynamic distribution shifts in both spatial and temporal dimensions, hence enhancing model performance.

Our extensive experiments on various real-world datasets and prediction backbones have demonstrated the robustness and generalisability of our method.