Research map / Label-efficient and robust learning / Domain adaptation and generalization

Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming

NeurIPS 2025 · Label-efficient and robust learning › Domain adaptation and generalization · approach: Conditional distribution alignment · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes GraphRTA, which performs dual reprogramming (gradient-based weight pruning of the GNN plus learned node-feature and edge XOR modifications of the target graph) with an extra unknown-class output and Beta-mixture entropy-based domain adversarial learning for unsupervised open-set graph domain adaptation.

Problem

Unsupervised open-set graph domain adaptation: classify target nodes into known source classes while recognizing target-only classes as unknown, without a manually tuned threshold and without source bias or domain shift issues.

Approach

Reprograms both the model (gradient-importance-based pruning of weights to reduce source bias) and the target graph (learnable feature perturbations and budget-constrained edge XOR edits), adds an extra unknown-class logit to remove the threshold, and uses a Beta mixture over output entropy to split target nodes into known/unknown groups for a three-domain adversarial loss.

Key results

• On citation datasets GraphRTA reaches 66.26% accuracy on ACMv9→Citationv1 versus 59.75% for the best baseline G2Pxy, and it reports the best H-score in most settings.
• On ogbn-arxiv Arxiv I→Arxiv II, GraphRTA achieves 47.70% accuracy and 50.79 H-score, above the best baseline GRADE (47.18 H-score).
• Ablation: removing model reprogramming or graph reprogramming lowers H-score on D→C (59.86 and 60.34 vs 62.33 for full model).

Measured effects reported in the paper

Surprising findings

• Several open-set baselines reach moderate accuracy but very low H-scores (e.g., OpenWGL 49.98% accuracy but 5.57 H-score on ACMv9→Citationv1), showing accuracy alone is misleading.
• Entropy-threshold variant of open-set detection is consistently worse than the proposed 3-domain mixture-based approach.

Limitations

• Graph reprogramming uses simple feature and XOR edge edits rather than more expressive structure learning; alternatives (SUBLIME, SLAPS) were tested only on two settings.
• Performance drops with larger mask ratios and gains are sensitive to hyperparameters such as the sparsity ratio rho and the layer count (best at 2 layers).
• Experiments use only a small number of known classes (3 for citation/WebKB, 30 for arxiv), and the ogbn-arxiv baseline DANCE runs out of memory.

Experimental setup

Datasets: DBLPv7 Citationv1 ACMv9 ogbn-arxiv Cornell Texas Wisconsin

Benchmarks: DBLPv7 Citationv1 ACMv9 ogbn-arxiv Cornell Texas Wisconsin

Metrics: accuracy H-score

Baselines: GCN GraphSAGE GAT UDAGCN GRADE SpecReg StruRW A2GNN DANCE OpenWGL PGL OpenWRF

Builds on: GCN domain adversarial learning lottery ticket hypothesis SUBLIME SLAPS

Closest papers by meaning

• Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.94)
• Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.92)
• Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.90)
• Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.89)
• Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.89)
• On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.88)