Research map / Label-efficient and robust learning / Domain adaptation and generalization
EvoluNet: Advancing Dynamic Non-IID Transfer Learning 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 EvoluNet, a framework and generalization bound for dynamic non-IID transfer learning on graphs, combining transformer-based multi-resolution temporal encoding with dual GRL-based domain unification to predict target graph labels at the next timestamp.
Problem
Transfer learning from label-rich dynamic source graphs to label-scarce dynamic target graphs, predicting node labels on the target graph at the incoming timestamp T+1 under evolving domain discrepancy.
Approach
Derives a generalization bound using the minimum historical empirical error and a dynamic Wasserstein graph discrepancy, and implements it as EvoluNet: a transformer temporal encoder with continuous-timestamp positional encoding plus spatial and temporal gradient-reversal domain-invariant losses.
Key results
- EvoluNet reaches AUC 0.6527 on Benchmark 1 (DBLP-5 to DBLP-3) vs 0.5823 for the best baseline EvolveGCN, a reported 12.1% improvement.
- On Benchmark 6 (DBLP-5 to HCP) EvoluNet achieves AUC 0.8046 vs 0.7407 for UDAGCN.
- Ablation: removing pre-training drops Benchmark 6 AUC from 0.8046 to 0.7234 (about 11.2%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add EvoluNet's dynamic temporal encoding and dual-divergence transfer framework | node classification on the six graph-transfer benchmarks · Benchmarks 1–6 | AUC: up to 12.1% | Helped |
| Add pre-training (ablation) | EvoluNet · Benchmark 1 | AUC: $0.5907$ -> $0.6527$ | Helped |
| Add pre-training (ablation) | EvoluNet · Benchmark 5 | AUC: $0.7661$ -> $0.7975$ | Helped |
| Add pre-training (ablation) | EvoluNet · Benchmark 6 | AUC: $0.7234$ -> $0.8046$ | Helped |
| Add module 1 (multi-resolution temporal encoding and attention) (ablation) | EvoluNet · Benchmark 1 | AUC: 0.6487 -> 0.6527 | Helped |
| Add module 1 (multi-resolution temporal encoding and attention) (ablation) | EvoluNet · Benchmark 5 | AUC: 0.7682 -> 0.7975 | Helped |
| Add module 1 (multi-resolution temporal encoding and attention) (ablation) | EvoluNet · Benchmark 6 | AUC: 0.7303 -> 0.8046 | Helped |
| Add spatial unification loss (ablation) | EvoluNet · Benchmark 1 | AUC: 0.6367 -> 0.6527 | Helped |
| Add spatial unification loss (ablation) | EvoluNet · Benchmark 5 | AUC: 0.7939 -> 0.7975 | Helped |
| Add spatial unification loss (ablation) | EvoluNet · Benchmark 6 | AUC: 0.7985 -> 0.8046 | Helped |
| Add temporal unification loss (ablation) | EvoluNet · Benchmark 1 | AUC: 0.6341 -> 0.6527 | Helped |
| Add temporal unification loss (ablation) | EvoluNet · Benchmark 5 | AUC: 0.7966 -> 0.7975 | Helped |
Surprising findings
- Temporal GNNs such as DCRNN, EvolveGCN and TGCN perform worse than static GCN on Benchmarks 5 and 6, failing to transfer.
- Removing the dual-divergence GRL losses barely changes or slightly improves results on Benchmarks 5 and 6, suggesting M2 contributes less there.
Limitations
- Benchmark is small (three graphs, six transfer pairs) and labels are assumed consistent across timestamps.
- Theoretical bound relies on Lipschitz and bounded Rademacher complexity assumptions and conditional independence of WL subtrees.
- M2 (domain-invariant GRL) shows inconsistent gains across benchmarks.
Experimental setup
Datasets: DBLP-3 DBLP-5 HCP
Benchmarks: DBLP-3 DBLP-5 HCP
Metrics: AUC
Baselines: GCN GAT GIN GraphSAGE DCRNN DyGrEncoder EvolveGCN TGCN DANN UDAGCN GRADE
Builds on: Wu et al. non-IID transfer learning on graphs (graph discrepancy) dynamic transfer learning bound DANN GRADE
Closest papers by meaning
- How Much Can Transfer? BRIDGE: Bounded Multi-Domain Graph Foundation Model with Generalization Guarantees (ICML 2025 · similarity 0.87)
- Temporal Generalization Estimation in Evolving Graphs (ICLR 2024 · similarity 0.87)
- Homophily Enhanced Graph Domain Adaptation (ICML 2025 · similarity 0.87)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.87)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.86)
- TMetaNet: Topological Meta-Learning Framework for Dynamic Link Prediction (ICML 2025 · 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.