Bridging the Fairness Divide: Achieving Group and Individual Fairness in Graph Neural Networks

1. Individual and group fairness are essentially different and probably contradicting objectives, unless the authors discuss why they can be compatible with each other. Even though the authors propose to consider both and try to achieve a better balance by enforcing "individual fairness within groups" instead of "individual fairness for all populations", this conflict could still confuse people and make the motivation less reasonable. The authors only make a statement in the "Challenges of balancing individual fairness and group fairness" paragraph of Section 4.2 while do not make sense to readers why this is reasonable. For example, wouldn't "individual fairness within groups" be essentially still different from "individual fairness" by ignoring individual fairness across groups? Post-hoc experimental analysis is currently not enough to answer this question. There is no explanation or theoretical insights why combining the two conflicting concepts can work well, but they are in fact at the core of supporting the motivation of this problem.
2. Compared to some existing works on fairness such as FairGNN, this work seems quite trivial since people can directly modify the loss objective of FairGNN (https://arxiv.org/pdf/2009.01454.pdf) or DebiAN (https://arxiv.org/pdf/2207.10077.pdf) to incorporate group fairness objectives. Unless authors can such modification to baselines in the comparison experiments and find out that even the modified version cannot compete with the proposed model, I can not convince myself why this work is giving a SOTA solution.
3. The motivation explanation with the toy example in Figure 1 seems strange. Why do group fairness objective select U2 and U3, instead of U3 and U4, since they both get SP=EO=0? All those numbers are only distances, but they do not indicate "score" or "grade". Seems to me U4 is qualified as well, so why not admit U4? Is there any reason that group fairness would tend to select U2 and U3?
4. Moreover, the language used in this paper can potentially stir public outrage: what does “qualified student” mean? I believe the authors do not mean that some students are not “qualified” because they are not "good" in the current system. What if it turns out that something like "a good family" is used as a positive feature in your embedding, since gender is the only considered sensitive factor? How to define “good”? The whole meaning of algorithmic fairness is to eliminate systematic bias and discrimination, and we cannot achieve that if the authors cannot make sense in the core example, regardless of whatever experimental results the paper has shown. It could be the "not even wrong" situation. The way this example puts makes the discussion quite sensitive, if not controversial. Most universities receive a lot of qualified admissions each year, and they never refer to any students as incompetent or unqualified as long as they pass the minimum requirement. At least for this example and for the explanation, I am not convinced and think it can be inappropriate. Please seriously reconsider this example in your paper.

1. The contribution appears incremental, primarily integrating individual fairness loss and group fairness loss from existing studies without significant innovation.
2. The definition and input of the $L_p$ function in Equation 9 are ambiguous, as both $Z$ and $f$ are presented as inputs. Clarification is needed regarding the role and relationship of these variables within the function.
3. The results are limited to the Pokec-n dataset. Given that Pokec-z is a similarly prevalent dataset, its omission is notable. The inclusion of Pokec-z results could enhance the robustness and generalizability of the findings.
4. The trade-offs between utility performance, group fairness, and individual fairness are not adequately explored. A Pareto frontier representation could provide valuable insights into the interplay between these dimensions.
5. There are typographical errors in Section 3.2.2; specifically, the second line features two instances of $V_{p_2}$ that seem out of place and likely constitute errors.

1. The design of the model is very similar to FairGNN[1] and the model is not innovative. FairGNN[1] includes 3 parts: a GNN classifier, a sensitive attribute classifier, an adversary layer, so I think FairGI just simply adds individual fairness and EO-related losses to FairGNN[1].
2. Lack of comparison with the latest group fairness methods in the experiment. I suggest the authors to consider comparing with the latest group fairness methods like FairVGNN[3] or EDITS[4].
3. The hyperparameters of the model are excessive, which poses a challenge for tuning. I found a total of 6 fairness-related hyperparameters, and the optimal hyperparameters are different for different datasets (Table 3), which requires a lot of time and computational resources for tuning the parameter when applying different GNN backbones as well as datasets.
4. The experiments lack sensitivity analysis experiments on hyperparameters. Although the experimental results show that FairGI can achieve good results, there are six hyperparameters related to fairness involved in the method (Table 3), but the whole experimental part does not have sensitivity analysis for any of them, which I think is necessary.

Reference

[1] Dai, E., & Wang, S. (2021, March). Say no to the discrimination: Learning fair graph neural networks with limited sensitive attribute information. In Proceedings of the 14th ACM International Conference on Web Search and Data Mining (pp. 680-688).

[2] Agarwal, Chirag, Himabindu Lakkaraju, and Marinka Zitnik. "Towards a unified framework for fair and stable graph representation learning." Uncertainty in Artificial Intelligence. PMLR, 2021.

[3] Wang, Y., Zhao, Y., Dong, Y., Chen, H., Li, J., & Derr, T. (2022). Improving Fairness in Graph Neural Networks via Mitigating Sensitive Attribute Leakage. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 1938–1948).

[4] Dong, Y., Liu, N., Jalaian, B., & Li, J. (2022, April). Edits: Modeling and mitigating data bias for graph neural networks. In Proceedings of the ACM Web Conference 2022 (pp. 1259-1269).

The paper appears to lack substantial original contribution, seemingly combining ideas from two prior works:

• The adversarial training framework for fairness is drawn from "Say No to Discrimination: Learning Fair Graph Neural Networks with Limited Sensitive Attribute Information."
• The computation of the similarity matrix of individuals to ensure individual fairness is inspired by "BeMap: Balanced Message Passing for Fair Graph Neural Network."

Given this synthesis of pre-existing concepts without significant novel insight or advancement, the paper may not meet the high standards of originality and innovation typically expected for acceptance at ICLR.