Research map / Label-efficient and robust learning / Domain adaptation and generalization
TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering
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 TFGDA, a semi-supervised graph domain adaptation framework that aligns persistent-homology topology between input and latent spaces (STSA), aligns features on a sphere via sliced-Wasserstein on great circles (SDA), and uses mutual-information robust clustering with perturbations (RNC) to handle label scarcity.
Problem
Semi-supervised graph domain adaptation for node classification where the source graph has very few labels (5%) and the target graph is fully unlabeled; existing methods align features but ignore graph structure and overfit under label scarcity.
Approach
Combines topology alignment via persistent homology between input graph and spherical latent features (on random-walk subgraphs), sphere-based sliced-Wasserstein feature alignment in place of adversarial training, and a mutual-information clustering objective under learned feature perturbations that avoids pseudo-labels.
Key results
- TFGDA achieves 81.0% Micro-F1 on A→C, the best among compared methods (SGDA 75.6%).
- On C→A, TFGDA beats SGDA by +7.3% Micro-F1 and +10.0% Macro-F1 (per paper text).
- Ablations show each of STSA, SDA, and RNC improves over the baseline TFGDA-S, and the combination yields the best results.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TFGDA instead of SGDA | Semi-supervised graph domain adaptation · C→A | Micro-F1: 66.3±0.68 -> 73.6±0.61 | Helped |
| Use TFGDA instead of SGDA | Semi-supervised graph domain adaptation · C→A | Macro-F1: 62.3±0.96 -> 72.3±0.94 | Helped |
| Add STSA structure alignment (ablation) | TFGDA-S · A→C | Micro-F1: 55.8±1.76 -> 72.6±0.55 | Helped |
| Add SDA sphere-guided domain alignment (ablation) | TFGDA-S · A→C | Micro-F1: 55.8±1.76 -> 75.8±0.38 | Helped |
| Add RNC robustness-guided node clustering (ablation) | TFGDA-S · A→C | Micro-F1: 55.8±1.76 -> 74.4±0.46 | Helped |
| Add STSA to SDA (ablation) | TFGDA-D · A→C | Micro-F1: 75.8±0.38 -> 78.9±0.59 | Helped |
| Add STSA to RNC (ablation) | TFGDA-R · A→C | Micro-F1: 74.4±0.46 -> 78.4±0.64 | Helped |
| Add SDA to RNC (ablation) | TFGDA-R · A→C | Micro-F1: 74.4±0.46 -> 79.2±0.41 | Helped |
| Add RNC to STSA and SDA (ablation) | TFGDA-TD · A→C | Micro-F1: 78.9±0.59 -> 81.0±0.34 | Helped |
| Add SDA to STSA and RNC (ablation) | TFGDA-TR · A→C | Micro-F1: 78.4±0.64 -> 81.0±0.34 | Helped |
| Add STSA to SDA and RNC (ablation) | TFGDA-DR · A→C | Micro-F1: 79.2±0.41 -> 81.0±0.34 | Helped |
| Replace RNC with conditional entropy minimization (CEM) (ablation) | TFGDA-S · A→C and A→D | Micro-F1 | Hurt |
Limitations
- Performance depends on subgraph size and number of sampled subgraphs for STSA.
- Large feature perturbation scale in RNC degrades performance.
- Evaluated only on three small citation graphs with a single 5-class setting.
Experimental setup
Datasets: ACMv9 Citationv1 DBLPv7
Benchmarks: ACMv9 Citationv1 DBLPv7
Metrics: Micro-F1 Macro-F1
Baselines: GCN GraphSAGE GAT GIN DANN CDAN DANNGCN CDANGCN UDA-GCN AdaGCN CoCo StruRW
Builds on: SGDA CDAN DANN AdaGCN UDA-GCN CoCo
Closest papers by meaning
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.90)
- 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)
- Pairwise Alignment Improves Graph Domain Adaptation (ICML 2024 · similarity 0.90)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.90)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · 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.