Fairness Aware Counterfactuals for Subgroups

Summary

The authors identify various aspects of fairness that require consideration when assessing recourse bias between subgroups. They propose FACTS (Fairness Aware Counterfactuals for Subgroups), in an attempt to lay foundations for a framework that can be used to audit subgroup fairness through counterfactual explanations (CEs). Experimental evaluation is relatively well thought out and thorough.

Strengths

1. The paper is easy to read, with a logical structure and natural flow. In most cases, relevant examples are provided where necessary. 2. I believe the issue being addressed is complex, and hard conclusions regarding the effectiveness of fairness tools should be made sparingly. However, the paper does an excellent job of analysing the problem at a deeper level than that which currently exists in the literature [1]. Explicit separation between the micro and macro viewpoint is well positioned. 3. While there are still gaps to be filled, this paper lays a strong foundation for future research into the use of counterfactuals in assessing recourse bias/unfairness. It is worth noting that [2] conducts similar analyses of the potential pitfalls of bias assessment via subgroup comparison, arriving at a similar conclusion to this paper that recourses should be compared 1 to 1 for more reliable results. [1] Michael Kearns, Seth Neel, Aaron Roth, and Zhiwei Steven Wu. Preventing fairness gerrymandering: Auditing and learning for subgroup fairness. ICML 2018. [2] Dan Ley, Saumitra Mishra, Daniele Magazzeni. GLOBE-CE: A Translation-Based Approach for Global Counterfactual Explanations. ICML 2023.

Weaknesses

I appreciate the thoroughness of the authors in tackling an issue as complex as this. There are a few important gaps that I believe may have evaded their attention, which I will detail below. 1. The assumption that while cost is not exactly quantifiable, a given action's cost is uniform across the input space, and can thus be compared between instances, does not always hold. One simple example would be changing salary by amount X. Individuals commanding higher initial salaries are likely to find this action easier. I would thus modify the "oblivious to the cost function" claim appropriately (though it is still a more than reasonable starting point for now). 2. Evaluation of fairness appears quite sensitive to the set of actions A chosen (specifically the fixed cost used/described in lines 290-302). Such a budget can often handle poorly the asymmetries between subgroup cost distributions for a given action. That said, multiple cost budgets are evaluated in the experiments, to provide a more complete overview. It's also worth considering the idea of scaling the cost of certain action directions, as in [2], and evaluating fairness accordingly. 3. In a similar vein, frequent itemsets are prone to resolution issues with numerical features. Moving beyond the apriori approach in [3] is wise, though frequent itemsets often do not uncover the minimal cost directions to flip predictions. Additionally, if the motivation behind fp-growth is primarily efficiency, it would be useful to have some analysis of this. 4. Discussion of when and how the fairness metrics can be manipulated (as above) is needed. Additionally, in practical settings, individuals are each in control of diverse actions, often unique to them. The effect of such actions (not included in but confounding with the feature space) is worth pointing out. 5. A large quantity of metrics are introduced. Table 1 caption could do with more clarification so readers don't have to study the text to understand the highlighted values. [3] Kaivalya Rawal and Himabindu Lakkaraju. Beyond individualized recourse: Interpretable and interactive summaries of actionable recourses. NeurIPS 2020.

Questions

The bins chosen for numerical features are rather large, and so may not effectively capture the minimum cost recourses between subgroups if the numerical features are influential, which might commonly be the case. This is one of the issues with frequent itemset mining, and something that [3] originally suffered with- do the authors have any proposed solutions? From my experience in the area, this is a fairly serious shortcoming, and something that prevented me from awarding this paper an even higher score.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

4 excellent

Limitations

Described above. Limitations associated with choosing parameters for fairness metrics and determining appropriate action spaces should be discussed in the revision. The research contributes valuable insights and provides a robust foundation for future work, though I would suggest to keep in mind the inherent complexity of the fairness problem, which is indeed emphasised in the paper. Overconfidence in the reliability of the proposed framework may not encourage an appropriately critical approach to this multifaceted issue i.e. conclusions drawn should be cautiously optimistic and mindful of the above limitations.

Summary

The paper presents a framework called FACTS (Fairness Aware Counterfactuals for Subgroups) for auditing subgroup fairness through counterfactual explanations. The authors aim to formulate different aspects of the difficulty individuals face in achieving recourse, either at the micro level (individuals within subgroups) or at the macro level (subgroups as a whole). They introduce new notions of subgroup fairness that are robust and provide an efficient, model-agnostic, and explainable framework for evaluating subgroup fairness. The authors demonstrate the advantages and wide applicability of their approach through an experimental evaluation using benchmark datasets.

Strengths

+ The paper addresses an important and timely topic in machine learning, namely fairness in decision-making processes. + The authors propose a novel framework, FACTS, for auditing subgroup fairness, which is model-agnostic and highly parameterizable. + The paper provides a thorough explanation of the different notions of subgroup fairness and their implications. + The experimental evaluation demonstrates the effectiveness and efficiency of the proposed approach on benchmark datasets. + The paper is well written and organized.

Weaknesses

There are a few weaknesses: - The paper could benefit from a more detailed description of the methodology used in the experimental evaluation. While the authors have provided an overview of the experimental setup and the data collection process, a more comprehensive explanation of the specific steps taken would greatly enhance the clarity and reproducibility of the study. For example, providing information about the specific criteria used to select the sources or the process of extracting and preprocessing the data would be useful for identifying the potential data generating bias. - In addition, providing more information on the statistical analysis performed on the experimental results would enhance the credibility of the findings. The authors mainly use Comparative Subgroup Counterfactuals for evaluating. However, it remains unclear whether the observations are significant, and how do they generalize to high-dimensional data as well. - The framework may face challenges in cases where the protected attributes are not well-defined or are subject to interpretation. How does FACTS perform when the underlying sensitive attributes are unknown (which is esp. true for many applications)?

Questions

Please refer to the weakness section for main questions. In addition to points raised above, there are a few questions remained: - How does the proposed framework handle the issue of defining a cost function for recourse? - Are there any computational or scalability limitations of the FACTS framework? - How does the performance of the FACTS framework compare to other state-of-the-art approaches for auditing subgroup fairness? - The proposed framework relies on the assumption that counterfactual explanations can accurately capture the underlying causes of bias and provide actionable insights. - The effectiveness of the FACTS framework may vary depending on the specific dataset and domain. - The framework may face challenges in cases where the protected attributes are not well-defined or are subject to interpretation.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

The main paper does not have a limitation statement, although the authors mentioned a few in the conclusion. There's no broader societal impact statement about the potential negative societal impact.

Summary

This paper considers the problem of fairness of machine learning based decisions. The setting is as follows: there is a set of features X in R^n where X_n denotes the demographic group, and a classifier h: X to {0,1} and we have access to a dataset D of individuals with h(X)=0. Each individual can take actions in an action space A to potentially change their classification decision to h(a(X))=1: this is recourse, and a is the counterfactual explanation. The paper considers how different actions to achieve recourse can be fair on either a micro or macro level. They first define multiple notions for effectiveness of actions to achieve recourse, that depend on a group, a cost budget or an effectiveness level. They then define and recall multiple definitions to define fairness of recourse (6 such definitions in section 2.3). They propose an algorithm "FACTS" that enables one to check for subgroups (based on features X) where there exist unfairness based on the 6 definitions and allows one to find the interventions to achieve needed effectiveness rates. They evaluate and showcase their algorithm on four different datasets.

Strengths

Originality: I think the way this paper thinks about recourse is novel in terms of the trade offs between effectiveness, cost and subgroups. Furthermore their algorithm is novel and intuitive way to find violations of fairness of recourse constraints. Quality: the experimental results seem sound (code is provided) and definitions are well discussed. Clarity: I enjoyed reading this paper, in particular, the introduction. However, there are some minor changes that can improve the exposition. Significance: I think the FACTS tool can serve as a nice way to audit algorithms for fairness of recourse. Moreover, the definitions in section 2.3 and the exposition in the introduction is very insightful for the community in algorithmic recourse.

Weaknesses

- there are two limitations of the algorithm: the form of the predicates (subgroups) and the form of the actions. In particular the subgroups are conjunctions of multiple features (does not allow arbitrary subgroups) and second actions are also conjunctions of feature values. This will lead to finding too many subgroups and too many actions, that could have instead been grouped under one subgroup or one action. Second, the algorithm only applies to categorical features and actions.

Questions

- can you address the point in the weaknesses section? - does the algorithm generalize to multiclass demographic groups and multiclass outcomes? - how can one find recourse for continuous features based on the provided algorithm? - the current algorithm lacks statistical tests to check if the subgroup found and the violations are statistically significant, how can one ensure that the violations found are in fact significant? - what is the runtime of the algorithm and how does it scale when features have very large spaces (i.e. suppose feature X_2 can take 10000 values) comments: - I really liked the introduction, but best to separate related work from the introduction. I would also encourage the authors to make the intro more succinct to avoid repeating the explanations in later sections. - the figures to show the recourse are not easy to read or attractive, I would suggest the authors spend some time improving the design of the CSC (figure 2 and figure 3) - sections 2.2 and 2.3 now read as a list of definitions without much continuity or story between the definitions and constraints.

Rating

7: Accept: Technically solid paper, with high impact on at least one sub-area, or moderate-to-high impact on more than one areas, with good-to-excellent evaluation, resources, reproducibility, and no unaddressed ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

limitations are adequately addressed.

Summary

In this paper, the authors explore the fairness of recourse in detail and distinguish between the micro and macro viewpoints. Moreover, they propose an efficient, interpretable, model-agnostic, highly parameterizable framework, called FACTS, to audit for fairness of recourse and provide an interpretable summary of its finding.

Strengths

- The authors claim that their work is the first that audits for fairness of recourse at the subpopulation level. - The results of their work are explainable and interpretable, which makes it more practical and easier to take action in order to mitigate the bias for different subgroups. - The authors provide a thorough analysis of various fairness metrics and their advantages and limitations to motivate their work. They also define various subgroup recourse fairness metrics and produce separate subgroup rankings per definition.

Weaknesses

- The experiments section could be improved by including more datasets and comparisons with other fairness metrics and definitions.

Questions

- In general, how realistic is it to be oblivious to the cost function and have an equal choice for recourse?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

- The experimental section provides limited insight as it only contains one dataset.

Summary

The paper proposes a framework (termed as FACTS) for analysing the recourse fairness of a machine learning model. The work introduces multiple metrics for quantifying recourse fairness - both at micro level (individuals considered separate) and macro level (individuals considered together). Some proposed recourse fairness metrics are argued to be unaffected by the underlying cost metric. Finally, the work demonstrates the proposed framework for a logistic regression model by training it on several datasets. All except one dataset results are present in the supplementary material.

Strengths

1. Analysing model fairness from the perspective of recourse explanations is an important research problem. There exists much work in analysing the fairness of model predictions, but less on analysing if the generated recourses are fair. 2. The paper is very well-written. The motivation, definitions, and metrics are all clearly explained. The authors cover the current metrics clearly, highlight their weaknesses, and then introduce their proposed metrics. 3. Bibliography is extensive and covers most of the recent papers on the related topics.

Weaknesses

1. Experiments section seems insufficient. Please see below for some examples a) The experiments section and the supplementary material presents outputs of the FACTS framework for different datasets, however, it is unclear how to evaluate the efficacy of the proposed framework. It would have been great if the authors had included experiments using toy datasets to create a biased model and then demonstrated that FACTS can uncover the recourse bias while the existing methods don’t/do it with poor coverage. b) I think the current experiments could have been more detailed for example by covering other model classes (e.g., tree-based, neural networks) to understand if the proposed metrics generalise to non-linear and non-differentiable models. c) FACTS is argued to be scalable, interpretable and highly parametrizable (Line 113). However, there are no experiments to support this. It would be great to understand interpretability through a user study for example. Similarly, analysing how FACTS scales to continuous datasets would be very helpful because as it uses an itemset miming algorithms which generally don't scale well with continuous datasets as feature binning creates a large search space [1]. 2. Line 119 - FACTS can explore systematically the feature space. I wonder if there are any guarantees about this? If the feature dimension is high and the dataset has continuous features then the search space becomes too large to scale well. [1] Dan Ley, Saumitra Mishra, Daniele Magazzeni. Global Counterfactual Explanations: Investigations, Implementations and Improvements. ICLR Workshop 2022.

Questions

1. It seems to create counterfactual summaries, FACTS mines conjunctions of predicates for all positive samples and then uses them to find counterfactuals. Hwoever, doesn't this approach limit the coverage (effectiveness) because if the positive samples are not diverse (i.e., the feature space is not well covered), we may not find counterfactuals for many individuals, resulting in poor coverage. 2. I think FACTS have many similarities in terms of metrics with [2]. For example, micro and macro vs finding a global direction and scaling it locally. Similarly, ECD vs coverage-cost profiles. I would like to hear authors comment on how their work differs from GLOBE-CE. 3. Some points needs clarification/correction 1. What is "horizontal intervention". lines 342 and line 70 2. I think the term "effectiveness" (Line 193) is same as the term "coverage" used in the context of GCEs? See Ref[3] 3. Further, set of actions A can be considered as diverse CFs. ECD profiles same as in GCEs? Micro vs Macro too in G~LOBE-CE 4. Line 154: I don’t think this claim is correct. AReS[3], GLOBE-CE[2] and Gupta et al. have tried to address this problem before using global (group-level) counterfactuals. 6. Line 161: It would be helpful to understand authors comments on how FACTS compares to other global explanations framework in terms of efficiency. [2] A Translation-Based Approach for Global Counterfactual Explanations. Dan Ley, Saumitra Mishra, Daniele Magazzeni. GLOBE-CE: CoRR abs/2305.17021 (2023) [3] Beyond Individualized Recourse: Interpretable and Interactive Summaries of Actionable Recourses. Kaivalya Rawal and Himabindu Lakkaraju. NeurIPS 2020

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

2 fair

Presentation

4 excellent

Contribution

2 fair

Limitations

The authors do not discuss any limitations of their work. However, it is important to note, given that FACTS is proposed as a tool to investigate recourse fairness, detailed experimentation and analysis would be needed before using the tool in the real-world applications.

Response to authors

I would like to thank the authors for their time spent in addressing my concerns. While the response is mostly satisfying, I maintain my score of 7 and vote for acceptance of this work.

Thank you for your rebuttal

I thank the authors for their efforts in providing a thorough rebuttal. The response addressed most of my concerns. For now, I tend to maintain my score and I look forward to the discussion with other reviewers.