SUCRe: Selective Uncertainty-Aware Contrastive Representation for Graph Transfer Learning

arXiv cs.GR details SUCRe, a selective uncertainty-aware contrastive representation method for graph transfer learning. The core idea is to stop treating all transferred graph knowledge as equally trustworthy and instead adapt only the samples that look reliable under structural and distribution shift.

The method adds structure-aware entropy-based matching discrepancy to model both feature uncertainty and structural coherence during cross-graph adaptation. It also uses a domain-aware semi-hard negative sampling strategy to build more informative contrastive sets while filtering out unreliable cross-domain relationships, which should matter to anyone trying to keep graph models from learning the wrong associations.

For developers working with graph-based ML, the practical appeal is twofold: less negative transfer and less wasted compute. The authors say experiments on graph transfer benchmarks show competitive performance with improved efficiency, which makes the approach interesting for large, sparse, or label-starved pipelines where every transferred example can skew results.