Research map / Label-efficient and robust learning / Domain adaptation and generalization
Open-Set Cross-Network Node Classification via Unknown-Excluded Adversarial Graph 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.
Proposes UAGA, a separate-adapt framework for open-set cross-network node classification that adversarially separates unknown nodes, then aligns only known-class target nodes to the source using negative domain adaptation coefficients for unknown nodes.
Problem
Cross-network node classification where the target network contains extra classes absent from the labeled source network (open-set setting), requiring known-class alignment without negative transfer from unknown-class nodes.
Approach
Uses a (K+1)-class neighborhood-aggregation GAT classifier trained adversarially (OSBP-style) to separate unknown nodes, then a GRL-based domain discriminator with positive coefficient for known-class nodes and negative coefficient for pseudo-unknown nodes, so that only known classes are aligned across networks.
Key results
- On C→D at openness 4/9, UAGA reaches OS* 74.00 and HS 56.41, versus 61.78 OS* and 43.85 HS for the best prior O-CNNC baseline SDA.
- On C→A, UAGA achieves AUC 83.41 versus 66.88 for the strongest baseline in the table (UDANE).
- Ablation on C→D shows removing the lambda=-1 negative coefficient drops HS from 56.4 to 42.8.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use UAGA instead of the strongest listed baseline | O-CNNC methods · C→D | OS: 57.09 -> 69.21 | Helped |
| use UAGA instead of the strongest listed baseline | O-CNNC methods · C→A | AUC: 66.88 -> 83.41 | Helped |
| add a separation stage before adaptation (ablation) | UAGA · C→D | OS: 64.3 -> 69.2 | Helped |
| add an adaptation stage after separation (ablation) | UAGA · C→D | OS: 55.0 -> 69.2 | Helped |
| assign λ=-1 to target unknown nodes (ablation) | UAGA · C→D | HS: 42.8 -> 56.4 | Helped |
| use neighborhood aggregation instead of an MLP classifier (ablation) | UAGA · C→D | OS: 59.6 -> 69.2 | Helped |
| add self-training (ablation) | UAGA · C→D | OS: 58.2 -> 69.2 | Helped |
| use 32 embedding dimensions (ablation) | UAGA · C→D | overall performance | Helped |
| increase GNN encoder depth (ablation) | UAGA · C→D | overall performance | Hurt |
| set target classification loss weight β=0.1 (ablation) | UAGA · C→D | overall performance | Helped |
| set pseudo-unknown set size R=2000 (ablation) | UAGA · C→D | overall performance | Helped |
Surprising findings
- Replacing the neighborhood-aggregation classifier with an MLP drops HS on C→D from 56.4 to 0.03, indicating the classifier design is critical for unknown detection.
- Closed-set CNNC methods achieve high OS* but low AUC, since whole-distribution alignment aligns unknown nodes with the source.
Limitations
- Positive and negative domain adaptation coefficients are fixed at 1/-1 rather than adaptive.
- Relies on pseudo-labels from clustering and classifier agreement, with a hyperparameter R for the pseudo-unknown set chosen from a small set.
- Evaluated only on three citation networks with keyword attributes and a single openness setting in the main table.
Experimental setup
Datasets: Citation-v1 DBLP-v4 ACM-v8
Benchmarks: Citation-v1 DBLP-v4 ACM-v8
Metrics: OS* OS AUC HS
Baselines: OSBP OMEGA OODGAT G2Pxy UDAGCN AdaGCN SGDA SDA UDANE
Builds on: OSBP Gradient reversal domain adaptation (DANN) UDAGCN graph attention network OMEGA
Closest papers by meaning
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.87)
- Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming (NeurIPS 2025 · similarity 0.86)
- Adversarial Robust Generalization of Graph Neural Networks (ICML 2025 · similarity 0.86)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.86)
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.85)
- Improving Distinguishability of Class for Graph Neural Networks (AAAI 2024 · similarity 0.85)
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.