Spectral Gap-Driven Coarsening for Dynamic Graph Neural Networks
Experimental results demonstrate that the coarsening technique significantly accelerates dynamic GNN training and inference without compromising predictive performance, offering a practical path toward scalable dynamic graph learning.