Spatio-temporal data is the foundation of many intelligent applications, revealing the causal relationships between measurements at a specific location and historical data from the same or other locations. In this context, Adaptive Spatio-Temporal Graph Neural Networks (ASTGNNs) have emerged as a powerful tool for modeling these dependencies, particularly through data-driven approaches rather than predefined spatial graphs. Although this method offers higher accuracy, it also introduces greater computational demands. To address this challenge, this paper delves into the concept of localization within ASTGNNs and proposes an innovative perspective that spatial dependencies should dynamically evolve over time. We introduce DynAGS, a localized ASTGNN framework aimed at maximizing efficiency and accuracy in distributed deployments. This framework integrates dynamic localization, time-varying spatial graphs, and personalized localization, all coordinated around a lightweight central module known as the dynamic graph generator, which employs a cross-attention mechanism. The central module can integrate historical information in a node-independent manner, enhancing the feature representation of nodes at the current moment. Subsequently, the improved feature representation is used to generate dynamic sparse graphs without incurring expensive data exchanges while supporting personalized localization. Through performance evaluations on two core ASTGNN architectures and nine real-world datasets from different application domains, the results indicate that DynAGS surpasses existing benchmark methods, highlighting that dynamic modeling of spatio-temporal dependencies can significantly enhance the expressiveness, flexibility, and system efficiency of models, especially in distributed environments.
For convenient access, visit the link below or click “Read the Original” at the bottom
https://www.zhuanzhi.ai/vip/3c7e162c18d5d353bb4b72f6d2f80c2a
Click “Read the Original” to view and download this article