DGOTTA Adapts GNNs to Dynamic Graphs with Temporal Memory

Graph Neural Networks (GNNs) trained on static graphs often struggle when deployed on dynamic graphs, where both connections and node features continuously change over time. This "distribution shift" can severely degrade model performance during test-time inference. Existing Test-Time Adaptation (TTA) methods primarily focus on static graphs, leaving a significant gap for dynamic scenarios. To address this, a new framework called DGOTTA (Dynamic Graph Online Test-Time Adaptation) has been introduced. DGOTTA is designed to effectively adapt pre-trained Dynamic GNNs (DGNNs) during live testing. It comprises three core modules: first, temporal-aware augmentation, which diversifies test dynamic graphs to handle complex temporal and spatial shifts; second, memory-aware model prediction, which actively mitigates catastrophic forgetting of previously learned patterns; and third, consistency-guided online adaptation, which ensures temporal alignment and smoothness in memory updates. Extensive experiments conducted on three real-world datasets and across four different DGNN backbones demonstrate that DGOTTA significantly enhances generalization capabilities. It shows robust improvements under various distribution shifts and across multiple model architectures, making DGNNs more reliable in evolving environments.