SCARCE

A label-free graph-level out-of-distribution generalization framework that learns distributionally robust graph representations through two-stage training, comprising VGAE warm-up and adversarial environment confusion, achieving implicit feature reweighting.