Research map / Graph learning / Graph self-supervised learning
Disentangled Graph Spectral 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 DGSDA, which disentangles attribute alignment (via embedding alignment) from topology alignment, replaced by aligning Bernstein-polynomial spectral GNN coefficients across source and target graphs for unsupervised graph domain adaptation.
Problem
Unsupervised graph domain adaptation (UGDA) struggles because node attributes and topology are entangled and topology alignment depends on a topology model and accurate target pseudo-labels.
Approach
Disentangle the graph distribution shift into attribute alignment (standard UDA embedding alignment) and model alignment, directly aligning BernNet Bernstein polynomial filter coefficients between source and target domains, instead of aligning topology; a coefficient-magnitude regularizer tightens a theoretical DA bound.
Key results
- DGSDA reaches 83.57% on A→C and outperforms the second-best baseline JHGDA by 6.33% on D→C per the authors' analysis.
- On the social network Blog task (B1→B2), DGSDA reaches 65.64% vs. 48.19% for the second-best baseline StruRW, a gain of over 17%.
- DGSDA is reported best on 3 of 4 social datasets and most traffic/citation tasks, but underperforms on Brazil→Europe.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DGSDA instead of the second-best baseline | DGSDA · D→C citation network adaptation | mean accuracy (%): +6.33% | Helped |
| Use DGSDA instead of the second-best baseline | DGSDA · Blog dataset | mean accuracy (%): over 17% | Helped |
| Add spectral model alignment loss Lalign (ablation) | DGSDA · citation networks | mean accuracy (%) | Helped |
| Add attribute alignment loss Lmmd (ablation) | DGSDA · citation networks | mean accuracy (%) | Helped |
| Add model alignment loss (ablation) | DGSDA · citation networks | mean accuracy (%) | Helped |
| Use DGSDA rather than baselines | DGSDA · airport traffic network adaptation | mean accuracy (%) | Mixed |
Surprising findings
- Aligning spectral filter coefficients directly, without any explicit topology alignment or final embedding alignment, outperforms topology-alignment methods such as StruRW and PairAlign.
- DGSDA underperforms baselines on the Brazil-to-Europe airport task, which the authors attribute to limited dataset scale causing overfitting to hubs.
Limitations
- Performance is weak on the small Brazil-Europe airport task, attributed to limited dataset scale.
- Theoretical bound relies on estimated Lipschitz constants and an assumed edge perturbation bound.
- Only node classification evaluated, with baselines reproduced from their released code and hyperparameters tuned per method.
Experimental setup
Datasets: ArnetMiner ACMv9 (A) Citationv1 (C) DBLPv7 (D) BlogCatalog Twitch-DE Twitch-EN Airport Brazil (B) Airport Europe (E) Airport USA (U)
Benchmarks: ArnetMiner citation networks (A/C/D) BlogCatalog and Twitch social networks Airport traffic networks (B/E/U)
Metrics: mean accuracy (%) standard deviation
Baselines: GCN GAT DANE UDAGCN AdaGCN StruRW JHGDA KBL PairAlign A2GNN
Builds on: BernNet UDAGCN StruRW PairAlign SpecReg
Closest papers by meaning
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.89)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.89)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.88)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.88)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.88)
- 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.