Research map / Graph learning / Graph self-supervised learning
Graph Data Selection for Domain Adaptation: A Model-Free Approach
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 GraDate, a model-free graph data selector that minimizes a Graph Dataset Distance (based on LinearFGW and optimal transport) between source training data and validation data to pick training graphs for graph domain adaptation.
Problem
Graph-level classification under covariate distribution shift (graph density and size) where source training data may contain harmful graphs; existing GDA methods are model-centric and costly, and model-free data valuation (LAVA) is designed for images.
Approach
Use Fused Gromov-Wasserstein (LinearFGW) distance between graphs and an OT-based dataset distance (GDD) that combines graph distance with label distance; iteratively reweight training samples via gradient steps and top-k sparsification (GREAT) to minimize GDD to the validation set, then train vanilla GNNs or GDA methods on the selected subset.
Key results
- On graph density shift with GCN at 10% selection, GraDate reaches 0.805 on IMDB-BINARY versus 0.737 for random and 0.620 for LAVA.
- Coupled with GRADE at 20% selection, GraDate's selected data improves GDA performance on several datasets (e.g., IMDB-MULTI 0.386 vs 0.112 random for GRADE; MSRC_21 0.778 vs 0.626 random).
- Vanilla GNNs trained on 20% GraDate-selected data reach or exceed full-data GNN accuracy on several datasets, e.g., GIN on ogbg-molhiv 0.644 vs 0.596 full.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Select 20% of source graphs with GraDate instead of LAVA | GCN on IMDB-BINARY under graph density shift · IMDB-BINARY | accuracy: 0.620 {${\pm}$ 0.000} -> 0.855 {${\pm}$ 0.024} | Helped |
| Select 20% of source graphs with GraDate instead of LAVA | GIN on MSRC_21 under graph density shift · MSRC_21 | accuracy: 0.851 {${\pm}$ 0.007} -> 0.860 {${\pm}$ 0.007} | No clear effect |
| Use GraDate-selected data instead of full training data | AdaGCN on IMDB-BINARY under graph density shift · IMDB-BINARY | accuracy: 0.808 {${\pm}$ 0.015} -> 0.817 {${\pm}$ 0.024} | Helped |
| Use GraDate-selected data instead of full training data | GRADE on MSRC_21 under graph density shift · MSRC_21 | accuracy: 0.804 {${\pm}$ 0.011} -> 0.778 {${\pm}$ 0.015} | No clear effect |
| Use GraDate-selected data instead of LAVA-selected data | off-the-shelf GDA methods across graph density shift settings · Six graph datasets | accuracy | Mixed |
| Increase the selection ratio | GraDate and LAVA selection methods with GNNs · Graph density shift | accuracy | Mixed |
Surprising findings
- Larger selection ratios do not always yield better performance for selection methods, since under severe shift a larger portion may contain irrelevant or harmful graphs.
- GraDate occasionally underperforms random selection when paired with GraphSAGE, attributed to neighbor sampling noise.
Limitations
- Performance can drop when LinearFGW approximation error leads to over-optimization on GDD (LAVA occasionally wins marginally).
- Experiments are limited to graph classification with covariate shifts (density and size) on six datasets; datasets lack node features in several cases.
- Label signal c requires tuning over a small set {0, 5}.
Experimental setup
Datasets: IMDB-BINARY IMDB-MULTI MSRC_21 ogbg-molbace ogbg-molbbbp ogbg-molhiv
Benchmarks: IMDB-BINARY IMDB-MULTI MSRC_21 ogbg-molbace ogbg-molbbbp ogbg-molhiv
Metrics: accuracy
Baselines: Random selection LAVA KiDD-LR AdaGCN GRADE ASN UDAGCN
Builds on: LinearFGW OTDD LAVA KiDD
Closest papers by meaning
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.89)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.89)
- Rethinking Propagation for Unsupervised Graph Domain Adaptation (AAAI 2024 · similarity 0.89)
- Domain Adaptive Unfolded Graph Neural Networks (AAAI 2025 · similarity 0.87)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.87)
- Pairwise Alignment Improves Graph Domain Adaptation (ICML 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.