Research map / Graph learning / Graph self-supervised learning
Homophily Enhanced 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 HGDA, a graph domain adaptation method that aligns homophilic, full-pass and heterophilic filtered signals across source and target graphs to reduce homophily distribution shift, with a PAC-Bayes bound motivating it.
Problem
Graph domain adaptation for cross-network node classification suffers when local node homophily distributions differ between source and target graphs, a factor prior GDA methods ignore.
Approach
Uses three GCN-style filters (homophilic A-tilde, full-pass I, heterophilic L-tilde) to extract separate signals, aligns each across domains via KL divergence losses, and trains with source cross-entropy plus target entropy, motivated by a PAC-Bayes bound that decomposes the target loss into homophily, attribute and heterophily shift terms.
Key results
- HGDA reports average improvements of 1.10% on Airport, 6.50% on ACM, 1.40% on Citation and 2.20% on Blog over the best baselines, and up to 8.49% accuracy gain on MAG.
- On A3→A4 HGDA reaches 0.718 accuracy versus 0.619 for PA (best baseline shown for that pair in the ACM table).
- Ablations show HGDA_L, HGDA_F and HGDA_H each perform differently across datasets, with the full combined HGDA best in most settings.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the full HGDA method instead of the strongest listed baseline | Cross-network node classification · Airport U→B | accuracy: 0.695 -> 0.721 | Helped |
| Use the full HGDA method instead of the strongest listed baseline | Cross-network node classification · ACM A3→A4 | accuracy: 0.619 -> 0.718 | Helped |
| Use the full HGDA method instead of the strongest listed baseline | Cross-network node classification · Citation A→C | accuracy: 0.814 -> 0.829 | Helped |
| Use the full HGDA method instead of the strongest listed baseline | Cross-network node classification · Twitch DE→EN | accuracy: 0.766 -> 0.781 | Helped |
| Use the full HGDA method instead of the strongest listed baseline | Cross-network node classification · MAG US→JP | accuracy: 0.474 -> 0.531 | Helped |
| Add the full-pass, heterophilic, and homophilic filters and alignment together (ablation) | HGDA on Airport U→B · Airport U→B | accuracy: 0.683 -> 0.721 | Helped |
| Add the full-pass, heterophilic, and homophilic filters and alignment together (ablation) | HGDA on Citation A→D · Citation A→D | accuracy: 0.756 -> 0.791 | Helped |
| Add the full-pass, heterophilic, and homophilic filters and alignment together (ablation) | HGDA on MAG US→CN · MAG US→CN | accuracy: 0.488 -> 0.510 | Helped |
| Use only the homophilic filter and omit the other filters (ablation) | HGDA on ACM A3→A4 · ACM A3→A4 | accuracy: 0.718 -> 0.709 | Hurt |
| Use only the full-pass filter and omit the other filters (ablation) | HGDA on ACM A3→A4 · ACM A3→A4 | accuracy: 0.718 -> 0.701 | Hurt |
| Use only the heterophilic filter and omit the other filters (ablation) | HGDA on ACM A3→A4 · ACM A3→A4 | accuracy: 0.718 -> 0.713 | Hurt |
| Use only the homophilic filter and omit the other filters (ablation) | HGDA on Airport U→B · Airport U→B | accuracy: 0.721 -> 0.683 | Hurt |
Surprising findings
- The theoretical bound's proof appears to replace the heterophily distribution term with the feature-based KL decomposition, which weakens the link between the theorem and the claimed homophily-specific bound.
- Several reported baselines (UDAGCN, EGI, JHGDA) run out of memory on the MAG datasets, so those comparisons are missing.
Limitations
- Complexity analysis claims O(N^2 d) and does not account for sparse operations, and the Airport/ACM tables have repeated or inconsistent entries.
- Evaluation relies on a two-layer GCN backbone and only accuracy as the metric.
- Theoretical bound relies on approximations (sums approximating expectations) and an assumed concentration inequality with unspecified constant C_0.
Experimental setup
Datasets: Airport DBLPv8 ACMv9 Citationv2 Twitch BlogCatalog ACM PSP ACM PAP MAG
Benchmarks: Airport ACM Citation Blog Twitch MAG
Metrics: accuracy
Baselines: GCN DANN DANE UDAGCN ASN EGI GRADE-N JHGDA SpecReg PA
Builds on: GCN DANN UDAGCN PA SpecReg
Closest papers by meaning
- Pairwise Alignment Improves Graph Domain Adaptation (ICML 2024 · similarity 0.92)
- On the Benefits of Attribute-Driven Graph Domain Adaptation (ICLR 2025 · similarity 0.91)
- Revisiting, Benchmarking and Understanding Unsupervised Graph Domain Adaptation (NeurIPS 2024 · similarity 0.91)
- Open-Set Graph Domain Adaptation via Separate Domain Alignment (AAAI 2024 · similarity 0.90)
- TFGDA: Exploring Topology and Feature Alignment in Semi-supervised Graph Domain Adaptation through Robust Clustering (NeurIPS 2024 · similarity 0.89)
- Smoothness Really Matters: A Simple Yet Effective Approach for Unsupervised Graph Domain Adaptation (AAAI 2025 · similarity 0.89)
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.