UFO: A Unified Flow-Oriented Framework for Robust Continual Graph Learning

Researchers have introduced UFO, a novel framework designed to address critical challenges in Continual Graph Learning (CGL). While CGL aims to model evolving graphs over time, existing methods often fail when faced with noisy supervision caused by annotation errors or adversarial corruption. This study highlights a new failure mode termed 'catastrophic remembering,' where models inadvertently reinforce corrupted knowledge across tasks. To mitigate this, the proposed Unified Flow-Oriented framework employs flow-based generative modeling to create replay representations, effectively reducing catastrophic forgetting without storing historical data. Additionally, UFO estimates instance-level reliability scores to distinguish clean nodes from noisy ones, thereby alleviating the impact of corrupted supervision. Extensive experiments conducted on four benchmark graph datasets under varying noise ratios demonstrate that UFO consistently outperforms current methods in both accuracy and forgetting metrics. This advancement significantly enhances the practical applicability of graph learning models in real-world scenarios involving dynamic and imperfect data. The associated code has been made available to support further research and implementation.