Orbit-Adaptive Zero-Shot Forecasting on Spatio-Temporal Graph

Abstract

Spatio-temporal graph forecasting is crucial for numerous real-world applications but often suffers from severe data scarcity in target domains, necessitating knowledge transfer from data-rich sources. In this work, we push this to the extreme by studying zero-shot spatio-temporal graph learning, where only the target graph topology is accessible and no temporal observations from the target domain are available at all. This strict setting renders state-of-the-art node-adaptive spatio-temporal graph neural networks ineffective, as their node-specific parameters cannot be inferred without target data. To bridge this gap, we propose \sysname, an Orbit-Adaptive Graph Neural Network framework. Our key insight is that a node's temporal behavior is not only determined by its own historical series but is also informed by its structural role within the graph, which we formally characterize via automorphism orbits—an equivalence relation capturing symmetric positions in the graph. By sharing adaptation mechanisms across nodes belonging to the same orbit, \sysname learns domain-invariant, structure-aware representations that enable prediction on completely unseen graphs. Extensive experiments on six datasets across traffic, air quality, and pandemic domains demonstrate that \sysname outperforms existing transfer and naive zero-shot baselines, establishing a new paradigm for structure-driven zero-shot forecasting.