One Identity, One Trajectory: Trajectory-Persistent Dynamic Graph Condensation for Efficient Temporal Graph Learning

Dynamic graphs are ubiquitous in real-world systems, yet their temporal evolution makes repeated training of dynamic graph neural networks (DGNNs) prohibitively expensive. Dynamic graph condensation aims to compress large-scale graph sequences into smaller yet informative synthetic sequences while preserving downstream utility. However, existing snapshot-wise methods condense each snapshot in isolation, discarding the cross-snapshot node identities that DGNNs fundamentally rely on to model temporal dependencies. Accordingly, we propose Trajectory-Persistent Dynamic Graph Condensation (TP-DGC), which shifts the condensation unit from isolated snapshots to node trajectories observed across the entire temporal window. Specifically, TP-DGC compresses structure-aware trajectory states from the real graph into a compact set of persistent synthetic identities. A time-shared assignment enforces index-consistent identities across all snapshots, making persistence intrinsic to the condensed representation rather than a post-hoc alignment. These identities are refined through trajectory-level distribution matching and task-specific semantic alignment. The resulting condensed trajectories are edge-free yet encode structural information through these persistent identities, enabling direct transfer across different DGNN architectures without re-condensation. Experiments on DBLP, Reddit, Brain, and ogbn-arxiv show that TP-DGC outperforms state-of-the-art baselines on node-classification across various condensation budgets, while achieving 7.1×–56.9× speedup over the strongest baseline in total condensation time.