PromptDyG: Test-Time Prompt Adaptation on Dynamic Graphs

Existing methods for Discrete-Time Dynamic Graphs (DTDGs) typically operate in an offline setting, where models are trained on historical snapshots and evaluated on future ones without further updates. This approach fails to account for the evolving nature of graph structures and the resulting distribution shifts between training and testing snapshots. The authors investigate how to adapt pre-trained dynamic graph models to these shifts in a live-update online setting without the computational burden of full model retraining.

The authors propose PromptDyG, a framework that introduces a lightweight, learnable prompt into the feature space of the test graph. During inference, the backbone model remains frozen to preserve the integrity of the learned temporal dynamics. The prompt is optimized at each time step using unsupervised feature-wise entropy minimization. This process forces the model to produce more discriminative embeddings, effectively bridging the gap between the training distribution and the shifted test-time structure.

PromptDyG consistently outperforms state-of-the-art DTDG models and existing graph test-time adaptation methods across six benchmark datasets. Theoretical and empirical analysis shows that the prompt adaptation significantly increases the similarity margin between positive (existing) and negative (non-existing) node pairs. Furthermore, the framework acts as a versatile, plug-and-play module that can enhance the performance of various pre-existing DTDG architectures.

This work addresses a critical limitation in dynamic graph learning: the inability of static, pre-trained models to handle the continuous evolution of real-world systems. By utilizing prompt-based adaptation, the authors provide a computationally efficient way to maintain high predictive accuracy in dynamic environments where data distributions are constantly changing, making it highly relevant for applications like transaction monitoring and social network analysis.