Research map / Label-efficient and robust learning / Domain adaptation and generalization
Revisiting, Benchmarking and Understanding 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.
Presents GDABench, a benchmark of 16 unsupervised graph domain adaptation methods on five real-world datasets (74 adaptation pairs), plus SimGDA/SimGDA+ GNN-variant studies and the PyGDA library.
Problem
Existing unsupervised graph domain adaptation (UGDA) methods are evaluated under inconsistent settings, datasets, and splits, making it unclear which methods work best and when, and how GNN aggregation affects transferability under graph distribution shifts.
Approach
A unified benchmark and library covering 16 UGDA methods across datasets with characterized feature, structure, and label shifts, plus a study of GNN aggregation variants (SimGDA) and unsupervised add-ons (SimGDA+) showing simple GNNs with proper aggregation can match or beat specialized UGDA methods.
Key results
- SimGDA+ achieves the best Micro-F1 on Airport E→U (58.11) and U→E (57.52) among all compared methods in the paper's table.
- A2GNN attains the top Macro-F1 on MAG FR→JP (26.20) and JP→FR (25.78) and AUROC-style leading results on several tasks, while SimGDA+ is competitive across datasets.
- On LLM-enhanced ogbn-arxiv, word2vec-encoded LLM text reduces feature shift and improves most UGDA baselines, whereas BERT-encoded text increases shift and degrades performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add unsupervised graph learning techniques to SimGDA | SimGDA · Airport E→U and U→E; Blog B1→B2 | Micro-F1: 55.29 -> 58.11; 54.39 -> 57.52; 53.35 -> 57.04 | Helped |
| Add unsupervised graph learning techniques to SimGDA | SimGDA · ArnetMiner C→A and D→A; Twitch DE→ES | Micro-F1 / AUROC: 70.80 -> 73.18; 67.04 -> 71.81; 61.30 -> 61.53 | Mixed |
| Add unsupervised graph learning techniques to SimGDA | SimGDA · MAG FR→JP, JP→FR, and JP→RU | Macro-F1: 18.59 -> 21.94; 15.16 -> 21.36; 13.27 -> 15.64 | Helped |
| Replace original node features with arxiv-LLM-w2v features | Graph domain adaptation models on ogbn-arxiv · 1950-2016→2016-2018 and 1950-2016→2018-2020 | Micro-F1 | Mixed |
| Replace original node features with arxiv-LLM-bert features | Graph domain adaptation models on ogbn-arxiv · Three publication-year transfer tasks | Micro-F1 | Hurt |
| Remove neighbor information (ablation) | SimGDA GNN variants · ArnetMiner, Blog, and Airport | Performance | Hurt |
| Remove neighbor information (ablation) | SimGDA GNN variants · Twitch and MAG | Performance | Helped |
| Add multi-hop neighbors (ablation) | SimGDA GNN variants · Airport and Blog | Performance | Hurt |
| Add multi-hop neighbors (ablation) | SimGDA GNN variants · ArnetMiner, Twitch, and MAG | Performance | Helped |
| Replace the GCN sum aggregator with mean or max aggregation (ablation) | SimGDA GNN variants · Across evaluated tasks | Performance | Hurt |
| Replace the GCN aggregator with GAT or GIN aggregation (ablation) | SimGDA GNN variants · ArnetMiner, Blog, Airport, and Twitch | Performance | Hurt |
Surprising findings
- Simple GNN variants (SimGDA) with appropriate aggregation can match or surpass specialized state-of-the-art UGDA methods.
- Ignoring neighbours (MLP-like) improves performance on Twitch and MAG, which have heavy label shift.
Limitations
- Benchmark covers only node classification in the main study, with limited scope (source-free and open-set settings not covered).
- Some baselines (SAGDA) had no released code and were reproduced by the authors; DGDA ran out of memory on MAG.
Experimental setup
Datasets: Airport Blog ArnetMiner Twitch MAG ogbn-arxiv
Benchmarks: GDABench Airport Blog ArnetMiner Twitch MAG ogbn-arxiv
Metrics: Micro-F1 Macro-F1 AUROC MMD
Baselines: DANE ACDNE UDAGCN ASN AdaGCN CWGCN SAGDA DMGNN DGDA StruRW JHGDA KBL
Builds on: DANN MMD GCN GRADE A2GNN PairAlign
Closest papers by meaning
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.95)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.94)
- Towards Unsupervised Open-Set Graph Domain Adaptation via Dual Reprogramming (NeurIPS 2025 · similarity 0.92)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.91)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.91)
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.90)
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.