Towards Expanding-Node Spatial-Temporal Forecasting: A Structured Node Interaction Prompting Perspective

ICLR 2026 Conference Submission16184 Authors

Keywords: Spatial-temporal Series, Expanding-Node Forecasting, Node Prompting

Abstract: The rapid expansion of sensor systems, such as traffic networks, climate monitoring, and energy scheduling, poses new challenges for spatial-temporal series forecasting. While existing models have achieved strong performance under the fixed-node assumption, they rely on node-dependent parameters and fail to adapt when the network evolves, i.e., when old nodes are removed and new nodes with limited history are added. This expanding-node forecasting scenario introduces two critical challenges: (1) learning heterogeneous node representations without coupling learnable parameters to node count, and (2) enabling effective adaptation to new nodes with scarce observations. To tackle these challenges, we propose SNIP (Structured Node Interaction Prompting), a model-agnostic framework that constructs static spatial-temporal priors from historical observations and topology, and dynamically refines them during model training. Specifically, SNIP generates structured priors from three perspectives: periodic patterns across nodes, spatial-temporal interactions under time delays and graph structural information. These priors are projected into model as node promptings and then dynamically refined. For new nodes, SNIP initializes priors by similarity-weighted mixtures of old nodes and updates them with limited history, enabling efficient few-shot adaptation. Extensive experiments on multiple datasets demonstrate that SNIP outperforms state-of-the-art baselines in expanding-node scenarios. Beyond accuracy, SNIP provides plug-and-play generality and computational efficiency, bridging the gap between fixed-node precision and expanding-node adaptability in spatial-temporal forecasting.

Primary Area: learning on time series and dynamical systems

Submission Number: 16184