Research map / Graph learning / Graph self-supervised learning
Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised 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 TDSS, which applies Laplacian smoothing with random-walk neighbor sampling on the unlabeled target graph (plus MMD alignment and A2GNN classifier loss) to reduce structural shift in unsupervised graph domain adaptation, with a theoretical target-risk bound in terms of model smoothness.
Problem
Unsupervised graph domain adaptation suffers when structural shifts between source and target graphs cause large changes in GNN node embeddings; existing methods mostly align latent features and overlook structure.
Approach
Apply a Laplacian smoothness regularizer only on the target graph, over neighbors sampled by random walks (avoiding over-smoothing of k-hop sampling), and bound target risk by source risk, TVD, and target model smoothness.
Key results
- TDSS (random-walk variant) reaches average Micro-F1 0.7954 and Macro-F1 0.7906 across six tasks, versus 0.7795 / 0.7672 for A2GNN.
- TDSS improves GCN, GAT and SGC backbones on all six tasks when the smoothing constraint is added.
- On D→A, NMI of node clustering is 0.4175 for TDSS versus 0.3651 for A2GNN and 0.2911 for SpecReg.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add target-domain structural smoothing with random-walk sampling | A2GNN · Six cross-domain tasks (average) | Micro-F1: 3.05% improvement | Helped |
| Add target-domain structural smoothing with random-walk sampling | A2GNN · Six cross-domain tasks (average) | Macro-F1: 2.04% improvement | Helped |
| Replace random-walk sampling with k-hop sampling (ablation) | TDSS · Six cross-domain tasks | Micro-F1 | Mixed |
| Remove domain-alignment loss (set α = 0) (ablation) | TDSS · Most transfer tasks | F1 | Hurt |
| Remove smoothness regularization loss (set β = 0) (ablation) | TDSS · Transfer tasks | F1 | Hurt |
| Replace random-walk sampling with 1-hop neighborhood sampling (ablation) | TDSS · C → D and C → A | F1 | Mixed |
| Add smoothing constraint | GCN, GAT, and SGC · Six cross-domain tasks | Micro-F1 and Macro-F1 | Helped |
| Increase random-walk length λ (ablation) | TDSS · A → C | Micro-F1 | Hurt |
| Increase random-walk length λ (ablation) | TDSS · A → D | Micro-F1 | Mixed |
| Increase walk number γ to higher values (ablation) | TDSS · Transfer tasks | Micro-F1 | Hurt |
| Set domain-alignment weight α to 0.3 (ablation) | TDSS · Two analyzed transfer tasks | Micro-F1 | Helped |
| Increase α beyond 0.3 (ablation) | TDSS · Two analyzed transfer tasks | Micro-F1 | Hurt |
Surprising findings
- The k-hop sampling variant underperforms the random-walk variant, which the authors attribute to over-smoothing with larger hop counts.
Limitations
- Theoretical bound relies on assumptions such as compact support, bounded loss and a covering-number argument that is only sketched in the appendix.
- Gains over A2GNN are modest (about 2-3 points on average) and small on some tasks such as C→A for the k-hop variant.
- Evaluated only on three small citation graphs from ArnetMiner with a single shared feature space.
Experimental setup
Datasets: ACMv9 Citationv1 DBLPv7
Benchmarks: ACMv9-Citationv1-DBLPv7 transfer tasks (A→C, A→D, C→A, C→D, D→A, D→C)
Metrics: Micro-F1 Macro-F1 NMI
Baselines: DeepWalk node2vec ANRL GAT GraphSAGE SGC GCN CDNE AdaGCN ACDNE UDAGCN GRADE
Builds on: A2GNN MMD StruRW PairAlign SpecReg
Closest papers by meaning
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.94)
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.93)
- 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)
- Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming (NeurIPS 2025 · similarity 0.89)
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.89)
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.