Research map / Graph learning / Graph self-supervised learning
Reproducibility Study Of Learning Fair Graph Representations Via Automated Data Augmentations
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.
Reproduces the Graphair fair graph augmentation paper's node-classification claims (partially reproducing Claim 1) and extends it to fair link prediction on Cora, Citeseer and PubMed against FairAdj and FairDrop.
Problem
Verifying whether the claims of Graphair (automated graph augmentation for fair graph representation learning) hold in independent reproduction, and whether it generalizes from node classification to link prediction with fairness.
Approach
Re-implements Graphair from the DIG library with full seeding, uses a 10,000-epoch grid search for Pokec, and adapts Graphair to link prediction via Hadamard link embeddings and mixed/subgroup dyadic-level fairness metrics.
Key results
- On NBA, reproduced Graphair achieves ΔDP 1.31 ± 0.19 at 68.54% accuracy, similar to the original trade-off; Pokec results show better fairness but lower accuracy than original (e.g., Pokec-n 65.76% ACC, ΔDP 0.72).
- Ablations confirm feature masking and edge perturbation both reduce bias on NBA and Pokec-z, so Claim 2 is fully reproduced; Claim 3 (lower homophily and feature–sensitive correlation in fair view) is fully reproduced.
- In link prediction, Graphair has a comparable accuracy-ΔDP_m trade-off to baselines but a superior trade-off for subgroup ΔDP_s, e.g., Citeseer ΔDP_s 2.9 vs 65.6 to 68.9 for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| remove edge perturbation (ablation) | Graphair · NBA, Pokec-n, Pokec-z | accuracy: 68.54 ± 0.40 -> 72.68 ± 0.40 (NBA); 65.76 ± 0.02 -> 67.26 ± 0.16 (Pokec-n); 65.22 ± 0.02 -> 69.25 ± 0.10 (Pokec-z) | Helped |
| remove edge perturbation (ablation) | Graphair · NBA, Pokec-n, Pokec-z | ΔDP: 1.31 ± 0.27 -> 2.95 ± 1.12 (NBA); 0.72 ± 0.36 -> 2.11 ± 0.09 (Pokec-n); 1.32 ± 0.33 -> 6.56 ± 0.74 (Pokec-z) | Hurt |
| remove feature masking (ablation) | Graphair · NBA, Pokec-n, Pokec-z | ΔDP: 1.31 ± 0.27 -> 10.73 ± 1.12 (NBA); 0.72 ± 0.36 -> 3.83 ± 0.47 (Pokec-n); 1.32 ± 0.33 -> 5.19 ± 0.74 (Pokec-z) | Hurt |
| remove feature masking (ablation) | Graphair · NBA, Pokec-n, Pokec-z | accuracy: 68.54 ± 0.40 -> 67.79 ± 0.32 (NBA); 65.76 ± 0.02 -> 64.61 ± 0.36 (Pokec-n); 65.22 ± 0.02 -> 57.91 ± 0.13 (Pokec-z) | Mixed |
| increase Pokec training and grid-search epochs from 500 to 10,000 (ablation) | Graphair · Pokec-n and Pokec-z node classification | accuracy | Helped |
| use Graphair instead of baseline models | Graphair · Citeseer, Cora, PubMed link prediction | accuracy and AUC | Hurt |
| use Graphair instead of baseline models | Graphair · Citeseer, Cora, PubMed link prediction, subgroup fairness | ΔDP_s and ΔEO_s | Helped |
| use Graphair instead of baseline models | Graphair · Citeseer, Cora, PubMed link prediction, mixed fairness | fairness-accuracy trade-off | No clear effect |
| use Graphair instead of baseline models | Graphair · Citeseer, Cora, PubMed link prediction, subgroup fairness | fairness-accuracy trade-off | Helped |
| reproduce original Graphair experiment with the reproduced setup | Graphair · NBA node classification | accuracy-fairness trade-off | No clear effect |
| reproduce original Graphair experiment with the reproduced setup | Graphair · Pokec-n and Pokec-z node classification | accuracy-fairness trade-off | Hurt |
Surprising findings
- A 500-epoch search does not yield optimal Pokec results; performance and fairness improve with up to 10,000 epochs.
- Graphair's accuracy increased when feature masking was removed, deviating from the original paper's findings.
Limitations
- Pokec experiments required 10,000 epochs instead of the original 500, so the setup departs from the original paper.
- Claim 1 was only partially reproduced; Pokec accuracy-fairness trade-off did not match original baselines.
- Link prediction adaptation required modifications to the framework and fairness metrics.
Experimental setup
Datasets: NBA Pokec-n Pokec-z Cora Citeseer PubMed
Benchmarks: NBA Pokec-n Pokec-z Cora Citeseer PubMed
Metrics: accuracy AUC ΔDP ΔEO homophily Spearman correlation
Baselines: FairAdj FairDrop FairWalk GRACE GCA NIFTY FairAug
Builds on: Graphair DIG FairGraph FairAdj FairDrop
Closest papers by meaning
- Fair Graph U-Net: A Fair Graph Learning Framework Integrating Group and Individual Awareness (AAAI 2025 · similarity 0.84)
- FairGP: A Scalable and Fair Graph Transformer Using Graph Partitioning (AAAI 2025 · similarity 0.83)
- Equivalence is All: A Unified View for Self-supervised Graph Learning (ICML 2025 · similarity 0.82)
- Cross-View Graph Consistency Learning for Invariant Graph Representations (AAAI 2025 · similarity 0.82)
- DANCE: Dual Unbiased Expansion with Group-acquired Alignment for Out-of-distribution Graph Fairness Learning (ICML 2025 · similarity 0.82)
- Chasing Fairness in Graphs: A GNN Architecture Perspective (AAAI 2024 · similarity 0.82)
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.