Training on large-scale graphs has achieved remarkable results in graph
representation learning, but its cost and storage have attracted increasing
concerns. Existing graph condensation methods primarily focus on optimizing the
feature matrices of condensed graphs while overlooking the