Research map / Label-efficient and robust learning / Domain adaptation and generalization
Sparse Causal Discovery with Generative Intervention 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 SLOGAN, an unsupervised graph domain adaptation method combining sparse causal disentanglement (via IB/VIB objectives), generative cross-domain spurious-feature intervention, and category-adaptive pseudo-label thresholding, evaluated on graph classification benchmarks.
Problem
Unsupervised graph domain adaptation for graph classification under distribution shift, where causal and spurious features are entangled and global adversarial alignment discards useful substructures while leaving spurious couplings intact.
Approach
Disentangles graph representations into causal and spurious parts using mutual-information/variational bottleneck objectives, then swaps spurious features across source and target samples through a generative reconstruction model to break spurious couplings, with covariance constraints and class-adaptive confidence thresholds for pseudo-labels.
Key results
- SLOGAN averages 67.8% on PTC cross-dataset adaptation versus 65.5% for the prior best MTDF.
- On NCI1 SLOGAN averages 70.6% versus 69.5% for MTDF.
- On Letter-Med SLOGAN averages 73.5% versus 71.3% for MTDF; on TWITTER-Real-Graph-Partial it averages 64.7% versus 64.2%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SLOGAN instead of MTDF | UGDA · PTC (source→target, Avg.) | accuracy: 65.5 -> 67.8 | Helped |
| use SLOGAN instead of MTDF | UGDA · NCI1 (source→target, Avg.) | accuracy: 69.5 -> 70.6 | Helped |
| use SLOGAN instead of MTDF | UGDA · TWITTER-Real-Graph-Partial (source→target, Avg.) | accuracy: 64.2 -> 64.7 | Helped |
| use SLOGAN instead of MTDF | UGDA · Letter-Med (source→target, Avg.) | accuracy: 71.3 -> 73.5 | Helped |
| add unbiased discriminative learning (L_sup) (ablation) | SLOGAN · NCI1 (source→target, Avg.) | accuracy: 67.5 -> 70.6 | Helped |
| add invariant learning under intervention (L_inv) (ablation) | SLOGAN · NCI1 (source→target, Avg.) | accuracy: 69.5 -> 70.6 | Helped |
| add causal-spurious disentanglement (L_dis) (ablation) | SLOGAN · NCI1 (source→target, Avg.) | accuracy: 69.6 -> 70.6 | Helped |
Surprising findings
- Several general domain adaptation methods (e.g., DANN, ToAlign) underperform simple GNN baselines or InfoGraph on some splits, indicating global alignment can hurt graph transfer.
Limitations
- Theoretical bound relies on assumptions (stable causal mechanism across domains) and the proof is loosely argued.
- Improvements over the strongest baseline are modest (often under 1-2 points), and some cells are worse than baselines, e.g., NCI1 N0->N2 bounded by CoCo-style variance.
- Hyperparameter and component descriptions are underspecified (e.g., how the causal/spurious split and covariance constraint are implemented).
Experimental setup
Datasets: PTC NCI1 TWITTER-Real-Graph-Partial Letter-Med
Benchmarks: PTC NCI1 TWITTER-Real-Graph-Partial Letter-Med
Metrics: classification accuracy
Baselines: GIN GCN GAT GraphSAGE MeanTeacher InfoGraph DANN ToAlign DUA DARE-GRAM CoCo MTDF
Builds on: InfoNCE variational information bottleneck DANN CoCo MTDF
Closest papers by meaning
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.88)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.87)
- Quantifying Distributional Invariance in Causal Subgraph for IRM-Free Graph Generalization (NeurIPS 2025 · similarity 0.87)
- A Recipe for Causal Graph Regression: Confounding Effects Revisited (ICML 2025 · similarity 0.87)
- Out-of-Distribution Generalization on Graphs via Progressive Inference (AAAI 2025 · similarity 0.87)
- Graph Contrastive Invariant Learning from the Causal Perspective (AAAI 2024 · similarity 0.86)
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.