GEMFlow: Dynamic Graph Evolution-Aware Masked Pre-training for Traffic Flow Forecasting

ICLR 2026 Conference Submission18537 Authors

Keywords: Spatiotemporal Forecasting, Adaptive Graph Neural Network, Masked Autoencoder

Abstract: Traffic flow forecasting is inherently challenging due to the continuous evolution of spatial dependencies and the coexistence of heterogeneous temporal patterns. Most existing pre-training methods either rely on static graphs or employ generic masking strategies that overlook the dynamic nature of road networks, limiting their robustness and transferability. To overcome these limitations, we propose GEMFlow (Graph Evolution-aware Masking for traffic Flow forecasting), a novel pre-training framework that unifies masked representation learning with adaptive graph evolution modeling. Specifically, GEMFlow introduces a curriculum-style dynamic masking strategy that operates on temporal patches while conditioning the masking process on the evolution of graph structures. This design allows the model to emphasize informative temporal segments and adapt to structural drift across time, going beyond prior decoupled masking approaches. The learned graph evolution-aware representations can be seamlessly transferred to diverse downstream forecasting models without modifying their architectures. Extensive experiments on four real-world PeMS datasets demonstrate that GEMFlow achieves state-of-the-art performance, consistently improving accuracy, efficiency, and robustness. Moreover, qualitative analysis of the learned dynamic graphs reveals interpretable evolution patterns, highlighting the potential of GEMFlow as a versatile pre-training paradigm for spatiotemporal forecasting.

Primary Area: learning on time series and dynamical systems

Submission Number: 18537