Research map / Graph learning / Graph self-supervised learning
Domain Adaptive Unfolded Graph Neural Networks
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 cascaded propagation (CP), which feeds the output of an unfolded GNN's lower-level propagation back as its input, provably lowering the lower-level objective under domain shift, and combines it with MMD alignment for unsupervised graph domain adaptation.
Problem
Unsupervised graph domain adaptation (GDA) where GNNs trained on a labeled source graph degrade on an unlabeled target graph; the architectural role of GNNs in GDA has been largely overlooked.
Approach
Views unfolded GNNs (UGNNs) as bi-level optimization and observes the lower-level graph signal denoising objective increases under domain transfer; cascaded propagation re-runs the lower-level propagation on its own output, which is guaranteed not to increase that objective, and is applied architecturally on top of MMD alignment.
Key results
- APPNP_CP improves average Macro-F1 by 0.62% and Micro-F1 by 0.72% over the best baseline on citation GDA tasks; ElasticGNN_CP improves by 1.34% Macro-F1 and 0.94% Micro-F1.
- ElasticGNN_CP outperforms UDAGCN on EN→DE by 5.04% Macro-F1 and 4.39% Micro-F1.
- Adding CP on top of MMD raises ElasticGNN's A→C-style results further, e.g. D→A Macro-F1 from 74.49 to 75.74 and C→A from 73.16 to 76.72.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add cascaded propagation to APPNP, GPRGNN, or ElasticGNN alongside MMD (ablation) | UGNNs · Citation-network GDA tasks | Macro-F1 | Helped |
| Add MMD alignment loss (ablation) | APPNP, GPRGNN, and ElasticGNN · Citation-network GDA tasks | Macro-F1 | Helped |
| Set the MMD trade-off parameter ξ to 0 instead of using MMD (ablation) | ElasticGNN_CP · A→C | performance | Hurt |
| Increase ξ too far beyond the useful range (ablation) | ElasticGNN_CP · Citation-network GDA tasks | performance | Hurt |
| Apply cascaded propagation repeatedly (ablation) | UGNNs · Domain transfer | lower-level objective | Helped |
Surprising findings
- On the social network transfer task, APPNP_CP and GPRGNN_CP do not beat the best baseline (UDAGCN on DE→EN), and GPRGNN_CP underperforms A2GNN on EN→DE.
Limitations
- Repeated CP can degrade performance when the number of layers is high; absolute performance does not necessarily improve.
- Uses a generic MMD alignment loss that ignores graph-specific structure.
- CP and MMD are developed independently rather than jointly designed.
Experimental setup
Datasets: ACMv9 Citationv1 DBLPv7 Twitch Germany Twitch England
Benchmarks: ACMv9 Citationv1 DBLPv7 Twitch DE Twitch EN
Metrics: Macro-F1 Micro-F1
Baselines: DeepWalk Node2vec GCN GAT GIN UDAGCN ACDNE ASN AdaGCN GRADE SpecReg A2GNN
Builds on: APPNP GPRGNN ElasticGNN MMD
Closest papers by meaning
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.93)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.90)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.89)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.88)
- Unified Graph Neural Networks Pre-training for Multi-domain Graphs (AAAI 2025 · similarity 0.87)
- Graph Data Selection for Domain Adaptation: A Model-Free Approach (NeurIPS 2025 · similarity 0.87)
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.