SAOT Improves Continual Graph Learning by Preserving Structure

Self-supervised Continual Graph Learning (CGL) is a growing area of research focused on enabling models to learn sequentially from a stream of graph data without explicit label supervision. A common limitation of current CGL methods is their tendency to optimize individual nodes in isolation, which can lead to a progressive distortion of the overall relational structure within the graph as learning continues across different tasks. This structural degradation hinders the model's ability to maintain accurate inter-node correspondences. To address this, a new framework called SAOT (Structure-Aware Optimal Transport) has been developed. SAOT leverages optimal transport theory to explicitly capture and preserve the global relational structure within graph representations across sequential learning tasks. Additionally, it incorporates a mechanism for cross-task knowledge distillation, ensuring that previously acquired structural knowledge is retained. Extensive experiments on various CGL benchmarks demonstrate that SAOT significantly outperforms existing self-supervised baselines, achieving substantial accuracy gains, particularly in class-incremental learning scenarios.