SAOT framework improves continual graph learning by 15% with optimal transport

Self-supervised Continual Graph Learning (CGL) lets models learn from a sequence of graph tasks without labels, but existing methods rely on instance-level consistency objectives that treat each node independently. This fails to maintain the global relational structure between nodes, causing embeddings to drift as tasks accumulate. To address this, Yuting Zhang, Yanbei Liu, and colleagues from Tianjin University and Tianjin University of Technology propose SAOT (Structure-Aware Optimal Transport), a framework accepted at ICML 2026. SAOT leverages optimal transport — a mathematical tool for aligning distributions — to explicitly model and preserve inter-node correspondences across tasks. It also incorporates a cross-task knowledge distillation mechanism to retain previously learned structural patterns.

On four standard CGL benchmarks, SAOT significantly outperforms existing self-supervised approaches. In the Class-IL setting (tasks with disjoint classes), SAOT improves average accuracy by up to 5% on CoraFull-CL and by over 15% on Products-CL compared to state-of-the-art methods. These results demonstrate that preserving global graph structure via optimal transport is a more effective strategy than instance-level consistency alone. The work opens up new directions for continual learning on graph data, particularly in scenarios where labeled data is scarce and the graph's relational structure is critical for downstream tasks.