Pith. sign in

REVIEW 4 major objections 5 minor 72 references

Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing

T0 review · 4 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Graph unlearning introduces bias; a bi-level debiasing method removes it while preserving privacy and accuracy.

desk verdict Useful first pass at fair graph unlearning, but the global debiasing objective and the deployed aggregation may be optimizing different predictors; needs a revision that addresses this before the fairness claims can be trusted. read the letter →

arxiv 2505.09702 v1 pith:AEH2VDLK submitted 2025-05-14 cs.LG

Yezi Liu , Prathyush Poduval , Wenjun Huang , Yang Ni , Hanning Chen , Mohsen Imani This is my paper · ORCID
classification cs.LG
keywords graphunlearninggroupfairnessdemographicparityequalopportunityneuralnetworksshard-basedmembershipinferencenodeclassification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

Deleting a user's node or edge from a trained graph model can quietly change how the model treats protected groups, so the very act of complying with a deletion request can make otherwise acceptable predictions unfair. The paper establishes this empirically and then claims a fix: FGU, a graph unlearning method that first debiases each shard model locally and then aligns all shard models against a global fairness objective. The intended payoff is that after unlearning, demographic parity and equal opportunity gaps drop below those of ordinary retraining, accuracy stays near retraining levels, and membership-inference attacks perform no better than random guessing. In short, the paper aims to show that forgetting and fairness are compatible, at a computational cost far below full retraining.

What carries the argument

The load-bearing mechanism is the bi-level debiasing objective over a sharded model. FGU partitions the graph into $K$ shards, trains a shard model $\theta_k$ on each, and aggregates predictions by the weighted sum of shard posteriors with learned importance weights $\lambda_k$; for the fairness objective it instead aggregates parameter weights as $\tilde{\theta} = \sum_k \lambda_k \theta_k$. Shard-level debiasing adds a local demographic-parity penalty $F_k$ to each shard's retraining loss, while global alignment minimizes $L_{\mathrm{global}} = U_{\mathrm{global}} + \alpha F_{\mathrm{global}}$, where $F_{\mathrm{global}}$ measures the demographic-parity gap of the aggregated model over the whole remaining graph. The two losses are coupled through the shard objective $L_k = U_k + \alpha_k F_k + \beta_k L_{\mathrm{global}}$, and an alternating optimization updates shard weights every epoch and importance weights every $t_1$ epochs. This lets fairness be enforced at the level where bias actually enters, the message-passing aggregation of shard predictions, without retraining the entire dataset.

What would settle it

Run FGU on a fixed unlearning request with the global-alignment term disabled but the learned $\lambda$ kept, and compare $\Delta_{\mathrm{DP}}$ of the deployed posterior-averaged predictor; if the gap stays essentially unchanged, the global alignment is not carrying the reported fairness improvement. A complementary check is to evaluate $\Delta_{\mathrm{DP}}$ for predictions made from the parameter-averaged weights versus from the posterior average on the same nodes; a large discrepancy would show the objective is optimizing a predictor different from the one users see.

Watch

Extended reading notes

Core claim

The paper's central claim is that standard graph unlearning methods systematically introduce bias: after nodes or edges are removed, the post-unlearning model's predictions become strongly correlated with the sensitive attribute, and the gap grows with the deletion ratio and is worse when deletions come from the unprivileged group. FGU counters this with bi-level debiasing: within each shard it retrains with a demographic-parity regularizer, and across shards it aligns the aggregated model by a global loss that penalizes disparity in prediction rates between sensitive groups. On six datasets and across node and edge deletion requests, FGU reports $\Delta_{\mathrm{DP}}$ and $\Delta_{\mathrm{EO}}$ values below both retraining and all graph-unlearning baselines, accuracy and F1 comparable to fair retraining, membership-inference attack accuracy near 50%, and a better accuracy-fairness trade-off than fairness-aware GNN baselines trained directly on the remaining data. The paper positions FGU as the first graph unlearning framework that simultaneously preserves privacy of deleted data and fairness of the post-unlearning model.

Load-bearing premise

The fairness adjustment is computed on an average of the models' internal weights, while the predictions users actually see come from averaging the models' probability outputs; for a deep graph network these are different predictors, and the paper does not show that making one fair makes the other fair.

Editorial extensions

If this is right

  • Unlearning requests are not fairness-neutral: deleting nodes or edges shifts prediction rates across sensitive groups, so any deployment of the right to be forgotten on graph data should audit fairness before and after deletion.
  • Because FGU's fairness gains come from shard-level retraining plus global alignment on the remaining graph, the method needs no access to the original full dataset after partitioning, which keeps unlearning efficient.
  • If FGU is correct, fairness and deletion privacy do not trade off against each other: membership-inference attack accuracy stays at chance while demographic parity and equal opportunity gaps remain near fair-retraining levels.
  • The accuracy-fairness comparison against fairness-aware GNN baselines implies that debiasing before or during unlearning, rather than after, is the right intervention point for post-deletion models.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A natural next experiment is to compute the global-alignment loss on the posterior-averaged predictor instead of the parameter-averaged one; if results match, the mechanism transfers, and if not, the simpler shard regularizer is the active ingredient.
  • The paper's observation that unlearning from the unprivileged group induces more bias suggests a testable corollary: bias introduced by deletion should scale with the level of homophily in the graph and with the concentration of deletion in one sensitive group, which could predict where fair unlearning is hardest.
  • The same bi-level design could be applied to federated graph learning, where clients are natural shards and global alignment plays the role of server-side aggregation, extending fair unlearning to settings where data cannot leave its owner.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. This paper studies whether graph unlearning introduces group bias and proposes FGU, a shard-based exact graph unlearning method with two debiasing levels: a local regularizer applied when retraining each shard, and a global alignment term that penalizes demographic parity gaps of a parameter-averaged shard ensemble. The authors report on six datasets that FGU achieves lower ΔDP and ΔEO than graph unlearning baselines and fair retraining while maintaining utility and membership-inference privacy, and they include ablations, hyperparameter sensitivity, fairness-aware GNN comparisons, and a link prediction extension.

Significance. If the claims are substantiated, FGU would be a practical contribution to graph unlearning with group fairness, with efficiency advantages from sharding. The paper is broad in scope: six datasets, multiple unlearning ratios and request types, privacy attacks, and comparisons to fairness-aware GNNs. The experiments are extensive and the idea of bi-level debiasing is intuitive. However, the central mechanism of the global alignment step is not validated as described, and the hyperparameter reporting is contradictory, so the current version does not yet establish that the global module provides the claimed benefit.

major comments (4)
  1. [Section III-C vs Section III-A] The global alignment loss is computed on the parameter-averaged model θ~ = Σ_k λ_k θ_k (Eqs. 8-9), while the deployed inference described in Section III-A averages the shard models' posterior vectors. For a nonlinear GCN, Σ_k λ_k softmax(f_{θ_k}(x)) differs from softmax(f_{Σ_k λ_k θ_k}(x)), so reducing the ΔDP of the parameter-averaged model need not reduce the ΔDP of the posterior-averaged model used in the reported numbers. The paper provides no bound, monotonicity argument, or ablation that evaluates both aggregation rules. As a result, the observed fairness improvements may be entirely attributable to the local regularizer F_k or to tuning λ on the evaluation metric. Please add an analysis or experiment that either justifies the transfer or defines F_global directly on the posterior-averaged predictor used at inference.
  2. [Section IV-F and Appendix I] The manuscript reports conflicting values for the fairness regularization weights. Appendix I states 'ultimately selecting α = 0.5 and β = 1 for the experiments,' whereas Section IV-F states 'we choose α = 3.0 and β = 1.5 to perform FGU.' In addition, Section IV-F says α and β are varied over {0.5, 1.5, 3.0, 5.0, 7.0} but the Figure 5 axes enumerate {0.001, 0.01, 0.1, 0.5, 1, 5, 10}. Since the accuracy-fairness trade-off is highly sensitive to these weights (Figure 5), please state unambiguously which values produced Table I and the other main results, and re-run or justify the reported numbers under the correct hyperparameters.
  3. [Table IV] The ablation does not support the stated conclusion that FGU is better than its two variants. In Table IV, Local Debiasing alone attains ACC 77.1 ± 2.2 and F1 89.8 ± 0.4 with ΔDP 3.1 ± 1.6 and ΔEO 3.2 ± 2.9, while full FGU attains ACC 66.8 ± 1.1, F1 77.1 ± 2.1, ΔDP 2.8 ± 0.2, and ΔEO 2.9 ± 0.3. Thus the global alignment term reduces the fairness gaps by only about 0.3 points while lowering accuracy by more than 10 points. This is the opposite of a favorable fairness-utility trade-off and weakens the claim that bi-level debiasing is beneficial. Also, the second ΔEO row in Table IV should presumably be labeled ΔDP.
  4. [Section II and III] Because ΔDP (and to a lesser extent ΔEO) is both the training objective in F_k and F_global (Eqs. 4-5) and the evaluation metric in Table I, the fairness improvements over non-fair baselines are to a significant degree a fitted outcome. This is a common limitation of in-processing fairness methods, but the paper should acknowledge it explicitly and, ideally, report at least one additional fairness measure or a held-out sensitive attribute to support the claim of 'superior fairness' beyond the optimized objective.
minor comments (5)
  1. [Section IV-A and Appendix I] The metric names are swapped in the text: 'Demographic Parity Difference △EO' and 'Equal Opportunity Difference △DP' appear where Eq. (1) defines ΔDP and ΔEO in the opposite way.
  2. [Equation (2)] Equation (2) mentions a regularization parameter γ that does not appear in the equation, and the notation U_k(θ_k) is used before its definition in Eq. (3); please rewrite for clarity.
  3. [Appendix C] The dataset description lists 'Pokec-z and Pokec-z' where the second subset should be 'Pokec-n'.
  4. [Section IV-I] The runtime bullet reads 'FGU is consistently less than 2.1 seconds and 5.2 seconds on German and seconds on Bail'; the sentence is incomplete, and the name 'Amenisac' is not defined in the manuscript.
  5. [Section IV-B, observation 4] The stated percentages for FGU's ΔDP relative to GEditor, GEraser, GDelete, and GIF (5%, 6.0%, 5.8%, and 4.7%) do not match the German row in Table I; for example, 1.7/27.1 is about 6.3%, not 5%.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: FGU openly trains on the DP objective it evaluates, and the cited self-work is not load-bearing; the global-alignment aggregation mismatch is a correctness gap, not a circular reduction.

full rationale

The central fairness claim is an in-processing optimization result rather than a prediction derived from inputs. FGU explicitly minimizes F_k (Eq. 4) and F_global (Eq. 5), both of which are the same ΔDP quantity used as an evaluation metric, and then reports lower ΔDP/ΔEO than baselines; this is standard empirical validation of a fairness-regularized training method, not a fitted parameter renamed as a prediction, because the evaluation split is separate from the optimized training objective and the comparison includes external baselines and fair retraining. The paper's self-citations ([24], [18], [50], [64]) are incidental background references and are not used to justify the core debiasing mechanism or to forbid alternatives. The most substantive concern is not circularity: the global alignment loss in Eqs. (6)-(9) is defined on the parameter-averaged model θ~ = Σλ_k θ_k, while Section III-A states that deployed inference averages the shard posterior vectors; for nonlinear GNNs these predictors differ, and Table IV's ablation does not isolate whether the reported fairness gain transfers to the deployed posterior-averaged model. That is a validity/reproducibility gap, and the inconsistent hyperparameter reporting (α=3.0/β=1.5 in Sec. IV-F vs α=0.5/β=1 in Appendix I) is a reporting issue, but neither makes the derivation equivalent to its inputs by construction. No self-definitional, fitted-input, imported-uniqueness, or ansatz-smuggling step was found, so the circularity score is low.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or conceptual entities; it relies on tuned scalar weights (alpha, beta, per-shard alpha_k/beta_k, lambda), shard partitioning, and two assumptions about how fairness transfers across aggregation schemes. The ledger shows most of the empirical success rests on fitted regularization strengths plus an unverified transfer assumption.

free parameters (4)
  • Fairness regularization weights alpha and beta = Contradictory: Section IV-A reports alpha=0.5, beta=1; Section IV-F selects alpha=3.0, beta=1.5
    Global and local fairness loss weights tuned on validation set; the paper reports two different settings, so the exact configuration behind the main tables is unclear.
  • Per-shard regularization weights alpha_k and beta_k = Tuned locally, values not enumerated
    Section IV-I states alpha_k and beta_k are tuned locally for each shard; no values or ranges are given, making exact replication difficult.
  • Shard importance scores lambda_k = Learned by optimization (Eq 10)
    The lambda weights are optimized to minimize L_global and directly affect the final predictor; they are fitted parameters of the method, not constants from theory.
  • Number of shards K = 20 in link-prediction appendix; not stated for main node classification experiments
    Shard count affects the bias-utility trade-off and the privacy profile; the main experiments do not report the value used.
assumptions (4)
  • domain assumption The DP and EO gaps measured after unlearning reflect algorithmic bias, not merely base-rate shift
    Section I and Section IV-B interpret increased delta DP and delta EO after non-uniform deletion as 'bias introduced by graph unlearning', but deleting data from one sensitive group changes the group prior; demographic parity will move even for a Bayes-optimal classifier.
  • ad hoc to paper Debiasing the parameter-averaged model transfers to the posterior-averaged inference model
    Eq (8)-(9) optimize theta~ = sum lambda_k theta_k, while Section III-A inference averages posteriors from shard models; no argument or experiment establishes that the two predictors behave identically.
  • domain assumption Shard models trained on disjoint partitions and aggregated can substitute for a full-graph model
    Inherited from SISA and GEraser, but graph partitioning removes cross-shard edges, changing message passing; the paper does not quantify this effect on utility or fairness.
  • domain assumption Alternating optimization of theta_k and lambda converges to a useful minimum
    The update scheme in Eq (10) and Algorithm 1 has no convergence or stationarity analysis; the paper assumes the iterative procedure reaches a model with low global DP.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing." pith.science (2026). https://pith.science/paper/AEH2VDLK

@misc{pith2026250509702,
  author       = {Pith},
  title        = {Pith review of: Enabling Group Fairness in Graph Unlearning via Bi-level Debiasing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AEH2VDLK}},
  note         = {Machine review of arXiv:2505.09702}
}
read the original abstract

Graph unlearning is a crucial approach for protecting user privacy by erasing the influence of user data on trained graph models. Recent developments in graph unlearning methods have primarily focused on maintaining model prediction performance while removing user information. However, we have observed that when user information is deleted from the model, the prediction distribution across different sensitive groups often changes. Furthermore, graph models are shown to be prone to amplifying biases, making the study of fairness in graph unlearning particularly important. This raises the question: Does graph unlearning actually introduce bias? Our findings indicate that the predictions of post-unlearning models become highly correlated with sensitive attributes, confirming the introduction of bias in the graph unlearning process. To address this issue, we propose a fair graph unlearning method, FGU. To guarantee privacy, FGU trains shard models on partitioned subgraphs, unlearns the requested data from the corresponding subgraphs, and retrains the shard models on the modified subgraphs. To ensure fairness, FGU employs a bi-level debiasing process: it first enables shard-level fairness by incorporating a fairness regularizer in the shard model retraining, and then achieves global-level fairness by aligning all shard models to minimize global disparity. Our experiments demonstrate that FGU achieves superior fairness while maintaining privacy and accuracy. Additionally, FGU is robust to diverse unlearning requests, ensuring fairness and utility performance across various data distributions.

Figures

Figures reproduced from arXiv: 2505.09702 by the authors.

Figure 1
Figure 1. Fairness performance of a two-layer GCNs before (left) and [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An illustration of FGU: shard debiasing achieves equitable predictions within each shard model, and global alignment reduces disparities among sensitive groups across shard models. a shard model Mk is trained on each shard graph Gk. The objective of the shard training process is defined as: min λ,θ E i∈V0 " Uk(θk) X K k=0 λkMk(Xi , Ni), Yi !# , (2) with Uk(θk) = X i∈Vk −[Yi log(Yˆ i) + (1 − Yi) log(1 − Yˆ i)], (3) w… view at source ↗
Figure 3
Figure 3. Accuracy of graph unlearning methods on three datasets, under [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Accuracy and △DP trade-off comparison on three datasets and five fairness-aware GNNs. Results located in the bottom-right corner are preferable. unlearning ratios. The widening accuracy gap may result from the increased introduction of bias as the unlearning ratio rise…
Figure 5
Figure 5. Figure 5: Accuracy and △DP trade-off on German, Credit, and Pokec-n. Results located in the bottom-right corner are preferable. parameter sensitivity and find a good trade-off of achieving high accuracy with low △DP , we train FGU on all datasets with various α values. More spec…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

72 extracted references · 55 canonical work pages

  1. [1]

    Privacy and artificial intelligence,

    J. Curzon, T. A. Kosa, R. Akalu, and K. El-Khatib, “Privacy and artificial intelligence,” TAI, vol. 2, no. 2, pp. 96–108, 2021

    work page 2021

  2. [2]

    An overview of artificial intelligence ethics,

    C. Huang, Z. Zhang, B. Mao, and X. Yao, “An overview of artificial intelligence ethics,” TAI, vol. 4, no. 4, pp. 799–819, 2022

    work page 2022

  3. [3]

    The right to be forgotten,

    J. Rosen, “The right to be forgotten,” Stan. L. Rev. Online, vol. 64, p. 88, 2011

    work page 2011

  4. [4]

    2018 reform of eu data protection rules,

    E. Commission, “2018 reform of eu data protection rules,” European Commission, 2018. [Online]. Available: https://ec.europa.eu/commission/sites/beta-political/files/ data-protection-factsheet-changes en.pdf

    work page 2018

  5. [5]

    California consumer privacy act (ccpa),

    C. DoJ, “California consumer privacy act (ccpa),” 2018. [Online]. Available: https://oag.ca.gov/privacy/ccpa

    work page 2018

  6. [6]

    Artificial intelligence across europe: A study on awareness, attitude and trust,

    T. Scantamburlo, A. Cort ´es, F. Foffano, C. Barru´e, V . Dis- tefano, L. Pham, and A. Fabris, “Artificial intelligence across europe: A study on awareness, attitude and trust,” TAI, 2024

    work page 2024

  7. [7]

    Inductive rep- resentation learning on large graphs,

    W. Hamilton, Z. Ying, and J. Leskovec, “Inductive rep- resentation learning on large graphs,” NeurIPS, vol. 30, 2017

    work page 2017

  8. [8]

    Scalable graph condensation with evolving capabilities,

    S. Gong, M. Hashemi, J. Ni, C. Yang, and W. Jin, “Scalable graph condensation with evolving capabilities,” arXiv preprint arXiv:2502.17614 , 2025

    arXiv 2025

Show all 72 references
  1. [9]

    Graph convolutional neural networks for web-scale recommender systems,

    R. Ying, R. He, K. Chen, P. Eksombatchai, W. L. Hamilton, and J. Leskovec, “Graph convolutional neural networks for web-scale recommender systems,” in KDD. ACM, 2018, pp. 974–983

    2018

  2. [10]

    Molecular generative graph neural networks for drug discovery,

    P. Bongini, M. Bianchini, and F. Scarselli, “Molecular generative graph neural networks for drug discovery,” Neurocomputing, vol. 450, pp. 242–252, 2021

    2021

  3. [11]

    A review of graph neural networks in epidemic modeling,

    Z. Liu, G. Wan, B. A. Prakash, M. S. Lau, and W. Jin, “A review of graph neural networks in epidemic modeling,” in KDD, 2024, pp. 6577–6587

    2024

  4. [12]

    Graph unlearning,

    M. Chen, Z. Zhang, T. Wang, M. Backes, M. Humbert, and Y . Zhang, “Graph unlearning,” in CCS, 2022, pp. 499–513

    2022

  5. [13]

    Gnndelete: A general strategy for unlearning in graph neural networks,

    J. Cheng, G. Dasoulas, H. He, C. Agarwal, and M. Zitnik, “Gnndelete: A general strategy for unlearning in graph neural networks,” arXiv preprint arXiv:2302.13406 , 2023

    2023 arXiv

  6. [14]

    Grapheditor: An efficient graph representation learning and unlearning approach,

    W. Cong and M. Mahdavi, “Grapheditor: An efficient graph representation learning and unlearning approach,” 2022

    2022

  7. [15]

    Com- bining neural networks with personalized pagerank for classification on graphs,

    J. Klicpera, A. Bojchevski, and S. G ¨unnemann, “Com- bining neural networks with personalized pagerank for classification on graphs,” in ICLR, 2019

    2019

  8. [16]

    A survey on bias and fairness in machine learning,

    N. Mehrabi, F. Morstatter, N. Saxena, K. Lerman, and A. Galstyan, “A survey on bias and fairness in machine learning,” CSUR, vol. 54, no. 6, pp. 1–35, 2021

    2021

  9. [17]

    Fairness through awareness,

    C. Dwork, M. Hardt, T. Pitassi, O. Reingold, and R. Zemel, “Fairness through awareness,” in ITCS, 2012, pp. 214–226

    2012

  10. [18]

    White admitted by stanford, black got rejections: Exploring racial stereotypes in text-to-image generation from a college admissions lens,

    Y . Liu, W. Huang, Y . Ni, H. Chen, and M. Imani, “White admitted by stanford, black got rejections: Exploring racial stereotypes in text-to-image generation from a college admissions lens,” in AAAI 2025 Workshop on Preventing and Detecting LLM Misinformation (PDLM)

    2025

  11. [19]

    Fairness in machine learning for health- care,

    M. A. Ahmad, A. Patel, C. Eckert, V . Kumar, and A. Teredesai, “Fairness in machine learning for health- care,” in KDD, 2020, pp. 3529–3530

    2020

  12. [20]

    Why is my classifier discriminatory?

    I. Chen, F. D. Johansson, and D. Sontag, “Why is my classifier discriminatory?” NeurIPS, vol. 31, 2018

    2018

  13. [21]

    Fairness in credit scoring: Assessment, implementation and profit implications,

    N. Kozodoi, J. Jacob, and S. Lessmann, “Fairness in credit scoring: Assessment, implementation and profit implications,” EJOR, vol. 297, no. 3, pp. 1083–1094, 2022

    2022

  14. [22]

    Fairness in criminal justice risk assessments: The state of the art,

    R. Berk, H. Heidari, S. Jabbari, M. Kearns, and A. Roth, “Fairness in criminal justice risk assessments: The state of the art,” Sociological Methods & Research , vol. 50, no. 1, pp. 3–44, 2021

    2021

  15. [23]

    Gif: A general graph unlearning strategy via influence function,

    J. Wu, Y . Yang, Y . Qian, Y . Sui, X. Wang, and X. He, “Gif: A general graph unlearning strategy via influence function,” in WWW, 2023, pp. 651–661

    2023

  16. [24]

    Fairgraph: Automated graph debiasing with gradient matching,

    Y . Liu, “Fairgraph: Automated graph debiasing with gradient matching,” in Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, 2023, pp. 4135–4139

    2023

  17. [25]

    Equality of oppor- tunity in supervised learning,

    M. Hardt, E. Price, and N. Srebro, “Equality of oppor- tunity in supervised learning,” NeurIPS, vol. 29, 2016

    2016

  18. [26]

    Fed- eraser: Enabling efficient client-level data removal from federated learning models,

    G. Liu, X. Ma, Y . Yang, C. Wang, and J. Liu, “Fed- eraser: Enabling efficient client-level data removal from federated learning models,” in IWQOS. IEEE, 2021, pp. 1–10

    2021

  19. [27]

    Federated un- learning with knowledge distillation,

    C. Wu, S. Zhu, and P. Mitra, “Federated un- learning with knowledge distillation,” arXiv preprint arXiv:2201.09441, 2022

    2022 arXiv

  20. [28]

    Inductive graph unlearning,

    C.-L. Wang, M. Huai, and D. Wang, “Inductive graph unlearning,” in USENIX Security, 2023, pp. 3205–3222. 10

    2023

  21. [29]

    Machine unlearning,

    L. Bourtoule, V . Chandrasekaran, C. A. Choquette-Choo, H. Jia, A. Travers, B. Zhang, D. Lie, and N. Papernot, “Machine unlearning,” in S&P. IEEE, 2021, pp. 141– 159

    2021

  22. [30]

    Safe: Machine unlearn- ing with shard graphs,

    Y . Dukler, B. Bowman, A. Achille, A. Golatkar, A. Swaminathan, and S. Soatto, “Safe: Machine unlearn- ing with shard graphs,” arXiv preprint arXiv:2304.13169, 2023

    2023 arXiv

  23. [31]

    Fairness beyond disparate treatment & dis- parate impact: Learning classification without disparate mistreatment,

    M. B. Zafar, I. Valera, M. Gomez Rodriguez, and K. P. Gummadi, “Fairness beyond disparate treatment & dis- parate impact: Learning classification without disparate mistreatment,” in WWW, 2017, pp. 1171–1180

    2017

  24. [32]

    Certifying and removing disparate impact,

    M. Feldman, S. A. Friedler, J. Moeller, C. Scheidegger, and S. Venkatasubramanian, “Certifying and removing disparate impact,” in KDD, 2015, pp. 259–268

    2015

  25. [33]

    Fair machine unlearning: Data removal while mitigating disparities,

    A. X. Oesterling, J. Ma, F. du Pin Calmon, and H. Lakkaraju, “Fair machine unlearning: Data removal while mitigating disparities,” 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:260202771

    2023

  26. [34]

    Fmp: Toward fair graph message passing against topology bias,

    Z. Jiang, X. Han, C. Fan, Z. Liu, N. Zou, A. Mostafavi, and X. Hu, “Fmp: Toward fair graph message passing against topology bias,” arXiv preprint arXiv:2202.04187, 2022

    2022 arXiv

  27. [35]

    Addressing shortcomings in fair graph learning datasets: Towards a new benchmark,

    X. Qian, Z. Guo, J. Li, H. Mao, B. Li, S. Wang, and Y . Ma, “Addressing shortcomings in fair graph learning datasets: Towards a new benchmark,” arXiv preprint arXiv:2403.06017, 2024

    2024 arXiv

  28. [36]

    Efficient model updates for approximate unlearning of graph-structured data,

    E. Chien, C. Pan, and O. Milenkovic, “Efficient model updates for approximate unlearning of graph-structured data,” in The Eleventh ICLR , 2022

    2022

  29. [37]

    Pri- vacy risk in machine learning: Analyzing the connection to overfitting,

    S. Yeom, I. Giacomelli, M. Fredrikson, and S. Jha, “Pri- vacy risk in machine learning: Analyzing the connection to overfitting,” in 2018 IEEE 31st computer security foundations symposium (CSF) . IEEE, 2018, pp. 268– 282

    2018

  30. [38]

    Machine unlearning: A survey,

    H. Xu, T. Zhu, L. Zhang, W. Zhou, and P. S. Yu, “Machine unlearning: A survey,” 2023

    2023

  31. [39]

    Kernel dependence regularizers and gaussian processes with applications to algorithmic fairness,

    Z. Li, A. Perez-Suay, G. Camps-Valls, and D. Sejdinovic, “Kernel dependence regularizers and gaussian processes with applications to algorithmic fairness,” arXiv preprint arXiv:1911.04322, 2019

    1911 arXiv

  32. [40]

    Towards a unified framework for fair and stable graph representation learning,

    C. Agarwal, H. Lakkaraju, and M. Zitnik, “Towards a unified framework for fair and stable graph representation learning,” in UAI 2021: UAI, 2021

    2021

  33. [41]

    Say no to the discrimination: Learning fair graph neural networks with limited sen- sitive attribute information,

    E. Dai and S. Wang, “Say no to the discrimination: Learning fair graph neural networks with limited sen- sitive attribute information,” in Proceedings of the 14th ACM WSDM, 2021, pp. 680–688

    2021

  34. [42]

    Edits: Modeling and mitigating data bias for graph neural networks,

    Y . Dong, N. Liu, B. Jalaian, and J. Li, “Edits: Modeling and mitigating data bias for graph neural networks,” in WWW, 2022, pp. 1259–1269

    2022

  35. [43]

    Stochastic methods for auc optimization subject to auc-based fairness constraints,

    Y . Yao, Q. Lin, and T. Yang, “Stochastic methods for auc optimization subject to auc-based fairness constraints,” in AISTATS. PMLR, 2023, pp. 10 324–10 342

    2023

  36. [44]

    Learning fair graph representations via automated data augmenta- tions,

    H. Ling, Z. Jiang, Y . Luo, S. Ji, and N. Zou, “Learning fair graph representations via automated data augmenta- tions,” in ICLR, 2022

    2022

  37. [45]

    Mitigating unwanted biases with adversarial learning,

    B. H. Zhang, B. Lemoine, and M. Mitchell, “Mitigating unwanted biases with adversarial learning,” in AIES, 2018, pp. 335–340

    2018

  38. [46]

    Fairness-aware classifier with prejudice remover regu- larizer,

    T. Kamishima, S. Akaho, H. Asoh, and J. Sakuma, “Fairness-aware classifier with prejudice remover regu- larizer,” in ECML PKDD. Springer, 2012

    2012

  39. [47]

    Learn- ing adversarially fair and transferable representations,

    D. Madras, E. Creager, T. Pitassi, and R. Zemel, “Learn- ing adversarially fair and transferable representations,” ICML, 2018

    2018

  40. [48]

    A comprehensive survey on trustworthy graph neural networks: Privacy, robustness, fairness, and explainability. arxiv e-prints,

    E. Dai, T. Zhao, H. Zhu, J. Xu, Z. Guo, H. Liu, J. Tang, and S. Wang, “A comprehensive survey on trustworthy graph neural networks: Privacy, robustness, fairness, and explainability. arxiv e-prints,” 2022

    2022

  41. [49]

    Semi-supervised classifica- tion with graph convolutional networks,

    T. N. Kipf and M. Welling, “Semi-supervised classifica- tion with graph convolutional networks,” arXiv preprint arXiv:1609.02907, 2016

    2016 arXiv

  42. [50]

    Tinygraph: joint feature and node condensation for graph neural networks,

    Y . Liu and Y . Shen, “Tinygraph: joint feature and node condensation for graph neural networks,” arXiv preprint arXiv:2407.08064, 2024

    2024 arXiv

  43. [51]

    Graph neural networks: A review of methods and applications,

    J. Zhou, G. Cui, S. Hu, Z. Zhang, C. Yang, Z. Liu, L. Wang, C. Li, and M. Sun, “Graph neural networks: A review of methods and applications,” AI open, vol. 1, pp. 57–81, 2020

    2020

  44. [52]

    Graph odes and beyond: A comprehensive sur- vey on integrating differential equations with graph neu- ral networks,

    Z. Liu, X. Wang, B. Wang, Z. Huang, C. Yang, and W. Jin, “Graph odes and beyond: A comprehensive sur- vey on integrating differential equations with graph neu- ral networks,” arXiv preprint arXiv:2503.23167 , 2025

    2025 arXiv

  45. [53]

    A review on machine unlearning,

    H. Zhang, T. Nakamura, T. Isohara, and K. Sakurai, “A review on machine unlearning,” SN Computer Science , vol. 4, no. 4, p. 337, 2023

    2023

  46. [54]

    A survey of machine unlearning,

    T. T. Nguyen, T. T. Huynh, P. L. Nguyen, A. W.-C. Liew, H. Yin, and Q. V . H. Nguyen, “A survey of machine unlearning,” arXiv preprint arXiv:2209.02299 , 2022

    2022 arXiv

  47. [55]

    Trustworthy machine learning and ar- tificial intelligence,

    K. R. Varshney, “Trustworthy machine learning and ar- tificial intelligence,” XRDS: Crossroads, The ACM Mag- azine for Students , vol. 25, no. 3, pp. 26–29, 2019

    2019

  48. [56]

    Strategic and business planning practices of fast growth family firms,

    N. Upton, E. J. Teal, and J. T. Felan, “Strategic and business planning practices of fast growth family firms,” Journal of small business management , vol. 39, no. 1, pp. 60–72, 2001

    2001

  49. [57]

    Charter of fundamental rights and general data protection regulation,

    “Charter of fundamental rights and general data protection regulation,” 2019. [Online]. Available: https: //europa.eu/eurobarometer/surveys/detail/2222

    2019

  50. [58]

    On the privacy risks of al- gorithmic fairness,

    H. Chang and R. Shokri, “On the privacy risks of al- gorithmic fairness,” in 2021 IEEE European Symposium on Security and Privacy (EuroS&P) . IEEE, 2021, pp. 292–303

    2021

  51. [59]

    Data privacy and trustworthy machine learning,

    M. Strobel and R. Shokri, “Data privacy and trustworthy machine learning,” IEEE Security & Privacy , vol. 20, no. 5, pp. 44–49, 2022

    2022

  52. [60]

    Fairsin: Achieving fairness in graph neural networks through sensitive in- formation neutralization,

    C. Yang, J. Liu, Y . Yan, and C. Shi, “Fairsin: Achieving fairness in graph neural networks through sensitive in- formation neutralization,” in AAAI, vol. 38, no. 8, 2024, pp. 9241–9249

    2024

  53. [61]

    Link prediction techniques, applications, and performance: A survey,

    A. Kumar, S. S. Singh, K. Singh, and B. Biswas, “Link prediction techniques, applications, and performance: A survey,” Physica A: Statistical Mechanics and its Appli- cations, vol. 553, p. 124289, 2020

    2020

  54. [62]

    Link prediction 11 in social networks: the state-of-the-art,

    P. Wang, B. Xu, Y . Wu, and X. Zhou, “Link prediction 11 in social networks: the state-of-the-art,” arXiv preprint arXiv:1411.5118, 2014

    2014 arXiv

  55. [63]

    Fairdrop: Biased edge dropout for enhancing fairness in graph representation learning,

    I. Spinelli, S. Scardapane, A. Hussain, and A. Uncini, “Fairdrop: Biased edge dropout for enhancing fairness in graph representation learning,” TAI, vol. 3, no. 3, pp. 344–354, 2021

    2021

  56. [64]

    Promoting fairness in link prediction with graph enhancement,

    Y . Liu, H. Chen, and M. Imani, “Promoting fairness in link prediction with graph enhancement,”Frontiers in Big Data, vol. 7, p. 1489306, 2024

    2024

  57. [65]

    On dyadic fairness: Exploring and mitigating bias in graph connections,

    P. Li, Y . Wang, H. Zhao, P. Hong, and H. Liu, “On dyadic fairness: Exploring and mitigating bias in graph connections,” in ICLR, 2021

    2021

  58. [66]

    Towards effective and general graph unlearning via mutual evolution,

    X. Li, Y . Zhao, Z. Wu, W. Zhang, R.-H. Li, and G. Wang, “Towards effective and general graph unlearning via mutual evolution,” in AAAI, vol. 38, no. 12, 2024, pp. 13 682–13 690

    2024

  59. [67]

    Automatic differentiation in pytorch,

    A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer, “Automatic differentiation in pytorch,” 2017

    2017

  60. [68]

    Adam: A method for stochastic optimization,

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimization,” arXiv preprint arXiv:1412.6980 , 2014

    2014 arXiv

  61. [69]

    Fast graph representa- tion learning with pytorch geometric,

    M. Fey and J. E. Lenssen, “Fast graph representa- tion learning with pytorch geometric,” arXiv preprint arXiv:1903.02428, 2019. APPENDIX A. Additional Preliminaries

    1903 arXiv

  62. [70]

    Details about Graph Neural Network: Here, we only introduce the design of GCN [49], which is one of the most popular GNN architectures [50, 51, 52]. More specifically, each layer of GCN can be written as: H(k) =σ( ˜AH(k−1)W(k)), (11) where H(k) denote the representations of al...

    1990

  63. [71]

    TABLE VIII: The utility and fairness performances of FGU on German with sensitive attribute gender after unlearning

    10% deletion on the privileged group: The results are shown in Table VIII. TABLE VIII: The utility and fairness performances of FGU on German with sensitive attribute gender after unlearning. We assess the performance of a 10% data unlearning on the privileged group. Retrain F...

  64. [72]

    TABLE IX: The utility and fairness performances of FGU on German with sensitive attribute gender after unlearning

    30% deletion on the privileged group: The results are shown in Table IX. TABLE IX: The utility and fairness performances of FGU on German with sensitive attribute gender after unlearning. We assess the performance of a 30% data unlearning on the privileged group. Retrain Fair ...

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.