Self-Supervised Continual Graph Learning in Adaptive Riemannian Spaces

Continual graph learning routinely finds its role in a variety of real-world applications where the data with different tasks come sequentially. Despite success prior works, it still faces great challenges. On one hand, existing methods work zero-curvature Euclidean space, and largely ignore fact that curvature varies over com- ing sequence. other continual learners literature rely on abundant labels, but labeling practice is particularly hard especially for continuously emerging graphs on-the-fly. To address aforementioned challenges, we propose to explore challenging yet practical problem, self-supervised adaptive Riemannian spaces. In this paper, novel Graph Learner (RieGrace). RieGrace, first design an Adaptive GCN (AdaRGCN), unified coupled neural adapter, so space shaped by learnt each graph. Then, present Label-free Lorentz Distillation approach, which create teacher-student AdaRGCN The student successively performs intra-distillation from itself inter-distillation teacher as consolidate knowledge without catastrophic forgetting. particular, theoretically grounded Generalized Projection contrastive distillation space. Extensive experiments benchmark datasets show superiority additionally, investigate how changes