Research map / Graph learning / Graph self-supervised learning
Pairwise Alignment Improves Graph Domain Adaptation
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 Pair-Align, a graph domain adaptation method that reweights source-graph edges via pairwise label-pair density ratios (addressing conditional structure shift) and reweights the classification loss by estimated label-shift weights, evaluated on MAG, HEP pileup, Arxiv, DBLP/ACM and synthetic CSBM data.
Problem
Graph domain adaptation for node classification under structure shift, where the joint distribution of adjacency and labels differs between source and target graphs (conditional structure shift and label shift), causing GNNs trained on source to generalize poorly to target.
Approach
Estimate pairwise (edge-type) label density ratios between source and target via a linear system over classifier soft-label edge statistics, use them as source edge weights so neighborhood label distributions match conditionally on the center label (CSS), and reweight the cross-entropy loss with label-shift ratios estimated by a similar confusion-matrix system (LS).
Key results
- On MAG, PA-BOTH reaches 40.06% on US→CN versus 31.58% for the best baseline StruRW, and 45.16% on CN→US versus 37.08% for StruRW.
- On Pileup, PA-CSS achieves 56.00 F1 on PU10→30 versus 52.61 for SpecReg; PA-LS is stronger in high-to-low PU directions such as PU50→10 (71.17 F1 vs 70.93 StruRW).
- On synthetic CSBM with CSS plus LS, PA-BOTH achieves 95.55% accuracy versus 87.01% for the best baseline IWDAN in the second setting.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add CSS edge reweighting (γ) to PA-LS (ablation) | Pair-Align · MAG datasets | accuracy | Helped |
| Add label-shift loss weighting (β) to PA-CSS (ablation) | Pair-Align · Pileup PU30→10 and high-PU→low-PU settings | F1 | Helped |
| Use label-shift weighting (β) alone instead of CSS edge reweighting (ablation) | Pair-Align · Pileup low-PU→high-PU settings | F1 | Mixed |
| Use only degree-shift alignment rather than neighbor-class alignment (ablation) | Pair-Align · Synthetic CSBM degree-shift scenarios | accuracy | No clear effect |
| Combine CSS and label-shift components instead of using either alone (ablation) | Pair-Align · Synthetic CSBM CSS and LS | accuracy | Helped |
| Combine CSS and label-shift components instead of using either alone (ablation) | Pair-Align · MAG datasets | accuracy | Helped |
| Combine CSS and label-shift components instead of using either alone (ablation) | Pair-Align · Arxiv and DBLP/ACM | accuracy | Mixed |
Surprising findings
- Marginal alignment baselines such as DANN and IWDAN often perform at or near ERM on MAG, and UDAGCN runs out of memory on the MAG graphs.
- Degree shift alone causes minimal performance decay for ERM under mean pooling, suggesting node degree shift has little impact.
Limitations
- Relies on an SBM-style edge conditional independence assumption and on classifier soft labels that may be inaccurate, motivating L2 regularization of weight estimates.
- Gains on Arxiv and DBLP/ACM are modest and some baselines are close, and PA-CSS is inconsistent in some HEP directions (e.g., PU30→10 and PU140→30 favor PA-BOTH or PA-LS).
Experimental setup
Datasets: MAG (regional subsets) Pileup mitigation HEP dataset Arxiv DBLP ACM synthetic CSBM
Benchmarks: MAG regional graphs Pileup mitigation Arxiv DBLP/ACM synthetic CSBM
Metrics: accuracy F1
Baselines: ERM DANN IWDAN UDAGCN StruRW SpecReg
Builds on: StruRW IWDAN DANN GraphSAGE contextual stochastic block model
Closest papers by meaning
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.92)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.90)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.90)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.90)
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.89)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.88)
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.