Prototype-Driven Dynamic Adaptation for Streaming Spatio-Temporal Graphs

Abstract

Spatio-Temporal Graph Neural Networks (STGNNs) often struggle under streaming Spatio-Temporal Out-of-Distribution (STOOD) shifts, where spatial and temporal patterns evolve over time. To address this, we propose \textbf{ProNet}, a prototype-driven dynamic adaptation framework that enhances the OOD robustness of STGNNs in streaming settings. ProNet features three key components: (1) an \textbf{Adaptive Prototype Memory (APM)} that maintains and updates representative spatio-temporal prototypes, (2) a \textbf{Pattern Alignment Module (PAM)} that aligns current inputs with stored prototypes for stable knowledge fusion, and (3) a \textbf{Dynamic Knowledge Distillation (DKD)} mechanism with adaptive temperature control to balance adaptation and retention. Extensive experiments on real-world streaming datasets demonstrate that ProNet consistently improves prediction accuracy and robustness across multiple STGNN backbones, offering a lightweight and plug-and-play solution for handling dynamic spatio-temporal shifts.