Spatial-Temporal-Decoupled Meta Learning for Dynamic Node Attribute Forecasting

ICLR 2026 Conference Submission18574 Authors

Keywords: Dynamic Graph Neural Networks, Spatiotemporal Forecasting, Meta-Learning

Abstract: Dynamic graph forecasting has become increasingly important in various domains, such as social networks and transportation systems. While dynamic graph neural networks (GNNs) have shown promise in predicting future node attributes, they often fail to capture the complex spatial-temporal interactions between nodes, limiting their performance. In this paper, we introduce STDMD, a novel dynamic GNN model that incorporates meta spatial-temporal decoupling to effectively capture both spatial and temporal dependencies in node attributes. By leveraging meta-learning, STDMD adapts to the evolving spatiotemporal patterns of node data, improving the accuracy and robustness of predictions. Specifically, our model dynamically refines spatial and temporal representations through an iterative meta-optimization process, allowing for more effective learning of dynamic node interactions. Furthermore, STDMD is designed to generalize across different dynamic graph structures, making it highly scalable and adaptable to realworld applications. Experimental results on real-world datasets demonstrate that STDMD outperforms state-of-the-art baselines, showcasing its ability to model dynamic node attributes with greater precision and robustness.

Primary Area: learning on time series and dynamical systems

Submission Number: 18574