Research map / Label-efficient and robust learning / Domain adaptation and generalization
Open-Set Graph Domain Adaptation via Separate Domain Alignment
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.
Introduces open-set graph domain adaptation for node classification and proposes Separate Domain Alignment (SDA), which adversarially aligns confident target nodes and clusters uncertain ones via neighbor center clustering.
Problem
Cross-network node classification under domain shift where the target graph contains unknown classes absent from the labeled source graph; existing graph domain adaptation methods assume a closed label space and suffer negative transfer.
Approach
Splits target nodes into certain and uncertain groups by prediction entropy; certain nodes are aligned to the source via adversarial domain alignment, while uncertain nodes are clustered via a neighbor center clustering loss that pulls them toward K-means cluster centers and source class centers.
Key results
- SDA reaches average Acc 60.73% and HS 53.55% across six open-set graph DA tasks, versus 55.81% Acc / 45.16% HS for ASN (best closed-set graph DA baseline).
- On D→A, the ablation shows full SDA improves HS from 38.70 (cross-entropy only) to 50.86.
- DANCE is slightly better than SDA on the D→A task in Acc, the only exception noted.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SDA instead of the strongest baseline for Acc | Open-set graph domain adaptation · A→D | Acc: 60.54 -> 61.60 | Helped |
| Use SDA instead of the strongest baseline for Acc | Open-set graph domain adaptation · D→A | Acc: 53.27 -> 51.36 | Hurt |
| Use SDA instead of the strongest baseline for Acc | Open-set graph domain adaptation · A→C | Acc: 63.23 -> 64.47 | Helped |
| Use SDA instead of the strongest baseline for Acc | Open-set graph domain adaptation · C→A | Acc: 60.44 -> 61.67 | Helped |
| Use SDA instead of the strongest baseline for Acc | Open-set graph domain adaptation · C→D | Acc: 64.29 -> 67.51 | Helped |
| Use SDA instead of the strongest baseline for Acc | Open-set graph domain adaptation · D→C | Acc: 57.62 -> 57.74 | Helped |
| Add neighbor center clustering to source cross-entropy training (ablation) | SDA · D→A | HS: 38.70 -> 45.67 | Helped |
| Add adversarial domain alignment to source cross-entropy training (ablation) | SDA · D→A | HS: 38.70 -> 48.34 | Helped |
| Add neighbor center clustering alongside adversarial domain alignment (ablation) | SDA · D→A | HS: 48.34 -> 50.86 | Helped |
| Vary the neighbor-center cluster count K (ablation) | SDA · D→A | HS | Helped |
| Increase the neighbor center clustering loss weight β (ablation) | SDA · D→A | Accu | Hurt |
Surprising findings
- Plain GCN and GraphSAGE approach or surpass the closed-set graph DA methods UDAGCN and ASN in h-score, which the authors attribute to negative transfer when target data contains unknown classes.
Limitations
- Evaluated only on three citation networks with a single 5-class label setup.
- Relies on a fixed entropy threshold log(|Cs|)/2 and a fixed number of clusters K for the uncertain group.
- DANCE is slightly better than SDA on the D→A task in Acc.
Experimental setup
Datasets: ACMv9 DBLPv7 Citationv1
Benchmarks: ACMv9 DBLPv7 Citationv1
Metrics: Acc Acc_known Acc_unknown h-score
Baselines: GCN GraphSAGE DANN CDAN OSBP DANCE UDAGCN ASN
Builds on: DANCE OSBP ASN UDAGCN PPMI-based GCN
Closest papers by meaning
- Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming (NeurIPS 2025 · similarity 0.94)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.90)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.90)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.90)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.90)
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.