Research map / Graph learning / Graph self-supervised learning
TRACI: A Data-centric Approach for Multi-Domain Generalization on Graphs
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 TRACI, a data-centric multi-domain generalization method for GNNs that combines topology-aware projected-gradient adversarial samples (entropy plus high-frequency structure term) with prototypical Mixup and contrastive alignment across source graphs.
Problem
Generalizing GNNs to unseen target graphs with distribution shifts (structural and nodal) when only multiple labeled source graphs are available and no target graph data can be used during training.
Approach
Generates adversarial samples by gradient ascent on an objective combining prediction entropy and a graph-structure (high-frequency perturbation) term, applied to both node attributes and hidden features, and aligns domains with multi-sample prototypical Mixup over class-graph Gaussian prototypes plus supervised contrastive loss.
Key results
- On the Citations benchmark TRACI reaches 72.95 Micro F1 on AC->D, 74.20 on AD->C and 67.01 on CD->A, above the best baseline in each setting.
- On the Proteins benchmark TRACI averages 52.18 accuracy across P1-P4, versus 46.62 for MARIO, the strongest baseline in the table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TRACI instead of FLOOD | Citations · AC→D | Micro F1: 67.26±0.50 -> 72.95±0.77 | Helped |
| Use TRACI instead of FLOOD | Citations · AC→D | Macro F1: 63.34±0.67 -> 69.25±1.22 | Helped |
| Use TRACI instead of MARIO | Citations · AD→C | Micro F1: 65.81±1.50 -> 74.20±0.68 | Helped |
| Use TRACI instead of MARIO | Citations · AD→C | Macro F1: 60.75±1.05 -> 69.08±1.07 | Helped |
| Use TRACI instead of MARIO | Citations · CD→A | Micro F1: 63.76±0.39 -> 67.01±1.04 | Helped |
| Use TRACI instead of GCN | Citations · CD→A | Macro F1: 59.24±2.00 -> 65.03±0.35 | Helped |
| Use TRACI instead of the strongest baseline | Proteins · P1 | accuracy: 50.48 -> 51.58 | Helped |
| Use TRACI instead of the strongest baseline | Proteins · P2 | accuracy: 52.92 -> 58.65 | Helped |
| Use TRACI instead of the strongest baseline | Proteins · P3 | accuracy: 38.82 -> 49.73 | Helped |
| Use TRACI instead of the strongest baseline | Proteins · P4 | accuracy: 47.97 -> 48.78 | Helped |
| Use TRACI instead of MARIO | Proteins · Average | accuracy: 46.62 -> 52.18 | Helped |
| Add prototypical Mixup (ablation) | TRACI · Citations, AC→D / AD→C / CD→A | Micro F1 | Helped |
Limitations
- Theoretical analysis is limited to a binary classification setting with a VC-dimension bound and does not directly justify the full method.
- Evaluated only on node classification over two benchmarks with few source graphs.
Experimental setup
Datasets: ACMv9 Citationv1 DBLPv7 PDB protein single-chain graphs (four datasets)
Benchmarks: Citations benchmark (AC->D, AD->C, CD->A) Proteins benchmark (P1-P4)
Metrics: Micro F1 Macro F1 accuracy
Baselines: GCN GIN GAT SGC EGC ADA MAT FLOOD MARIO
Builds on: MAT FLOOD Mixup projected gradient descent adversarial training supervised contrastive learning
Closest papers by meaning
- GOttack: Universal Adversarial Attacks on Graph Neural Networks via Graph Orbits Learning (ICLR 2025 · similarity 0.89)
- Adversarial Robust Generalization of Graph Neural Networks (ICML 2025 · similarity 0.88)
- A Manifold Perspective on the Statistical Generalization of Graph Neural Networks (ICML 2025 · similarity 0.88)
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.88)
- Generalization of Graph Neural Networks Is Robust to Model Mismatch (AAAI 2025 · similarity 0.87)
- Unified Graph Neural Networks Pre-training for Multi-domain Graphs (AAAI 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.