SCARCE learns label-free graph representations robust to distribution shift
Original title: Implicit Feature Reweighting for Label-Free Graph Generalization
A two-stage autoencoder confuses environment signals while preserving reconstruction and leads label-free results on DrugOOD.
Researchers propose SCARCE, a label-free framework for graph-level out-of-distribution generalization. It uses an environment partition from metadata or from clustering a self-supervised embedding, without class labels during representation learning. A variational graph autoencoder is warm-started for reconstruction, then adversarial confusion hides environment identity while keeping reconstruction fidelity. On controlled synthetic data and DrugOOD, the authors report state-of-the-art results among label-free methods.
Unconfirmed: The state-of-the-art claim reflects the authors’ experiments, not an independent evaluation.
- Environment partitions may come from dataset metadata or clustering when metadata is absent.
- The theory says environment confusion contracts environment-discriminative representation directions.
- The method is framed as a continuous relaxation of last-layer feature reweighting.
- Class labels are not used while learning the graph representations.
From the original article
1 min read at arxiv.orgGraph-level out-of-distribution (OOD) generalization is critical for deploying graph neural networks in real-world applications where distribution shifts are inevitable. Existing approaches typically require class labels during representation learning to decide which substructures are predictive, limiting their applicability.
By Tianjun Yao · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.