Fairness in Machine Learning: A Survey

In the paper

“However, we are yet to ind a strategy using pre-, in-, and post-processing methods combined.”

In the paper

“While many researchers publish their individual approaches on GitHub and similar platforms, a few notable projects have emerged that address fairness in ML from a more general perspective.”

In the paper

“The fairness literature focuses on either the technical aspects of bias and fairness in ML, or theorises on the social, legal, and ethical aspects of ML discrimination [124].”

What the sources say

• Not checked yetcites [124]

In the paper

“data veracity and quality [320], data relativity and context shifts [42,122,249], and subjectivity ilters [41].”

What the sources say

• Not checked yetcites [320] source
• Not checked yetcites [42] source
• Not checked yetcites [122] source
• Not checked yetcites [249] source
• Not checked yetcites [41]

In the paper

“It is noteworthy that much of the early literature into fairness considered regression and correlation-based scenarios (standardized test scores, see e.g., [81]), compared to the recent emphasis on classiication (binary decisions).”

What the sources say

• Not checked yetcites [81] source

In the paper

“For example, discrimination in hiring decisions has been examined since the 1960s [133].”

What the sources say

• Not checked yetcites [133] source

In the paper

“Despite the increased discussion of diferent aspects and viewpoints of fairness in the 1970s as well as the founding of many modern fairness concepts, no general consensus as to what constitutes fairness or if it can/should be quantiied emerged based on this irst wave of fairness research.”

In the paper

“However, the general notions transfer to (binary) classiication settings and thus deine essential concepts such as protected/demographic variables (e.g.[72,73]), notions of group vs individual fairness (e.g.[257,278]), impossibility conditions between fairness conditions ( [81]), and metric-based fairness quantiication (e.g., true positive rates, [74]).”

What the sources say

• Not checked yetcites [72] source
• Not checked yetcites [73] source
• Not checked yetcites [257] source
• Not checked yetcites [278] source
• Not checked yetcites [81] source

In the paper

“The importance of fairness and related ethical principles, however, is recognized as key to improving the trustworthiness of ML (and AI in general) [75,198].”

What the sources say

• Not checked yetcites [75]
• Not checked yetcites [198]

In the paper

“Two initial foci of fairness research were unfairness of standardized tests in higher education contexts [72,73] and discrimination in employment contexts [133].”

What the sources say

• Not checked yetcites [72] source
• Not checked yetcites [73] source
• Not checked yetcites [133] source

In the paper

“Yet, for correlated or otherwise related variables (often referred to as proxies or quasiidentiiers), much of the literature assumes a priori knowledge of the set of quasi-identiiers [111], or seeks to discover them on a case-by-case basis (e.g.[150,217]) towards the idea of privacy preserving data mining [6].”

What the sources say

• Not checked yetcites [111] source
• Not checked yetcites [150] source
• Not checked yetcites [217]
• Not checked yetcites [6] source

In the paper

“Key approaches (e.g.[194,205,276]) seek to protect speciic individuals and groups from being identiiable, i.e., minimize disclosure risk.”

What the sources say

• Not checked yetcites [194] source
• Not checked yetcites [205] source
• Not checked yetcites [276] source

In the paper

“In particular, sensitive data disclosure is a long-standing challenge in protecting anonymity when data is published and/or analysed [6,111,195].”

What the sources say

• Not checked yetcites [6] source
• Not checked yetcites [111] source
• Not checked yetcites [195] source

In the paper

“Pre-processing is argued as the most lexible part of the data science pipeline, as it makes no assumptions with respect to the choice of subsequently applied modeling technique [88].”

What the sources say

• Not checked yetcites [88] source

In the paper

“All technical fairness interventions operate at speciic locations in the ML pipeline, i.e., before, during, or after model training.”

In the paper

“Thus, pre-processing approaches tend to alter the sample distributions of protected variables, or more generally perform speciic transformations on the data with the aim to remove discrimination from the training data [165].”

What the sources say

• Not checked yetcites [165] source

In the paper

“Yet, modiication of the data and/or model output may have legal implications [17] and can mean models are less interpretable [192,202], which may be at odds with current data protection legislation with respect to explainability.”

What the sources say

• Not checked yetcites [17]
• Not checked yetcites [192] source
• Not checked yetcites [202]

In the paper

“In-processing approaches tackle this by often incorporating one or more fairness metrics into the model optimization functions in a bid to converge towards a model parameterization that maximizes performance and fairness.”

In the paper

“Many are diicult / impossible to combine [69,177], but ultimately, we must keep in mind (as noted in [68]) there is no universal means to measure fairness, and no clear guideline(s) on which measures are łbestž.”

What the sources say

• Not checked yetcites [68]

In the paper

“However, increasing fairness often results in lower overall accuracy or related metrics, leading to the necessity of analyzing potentially achievable trade-ofs in a given scenario [135].”

What the sources say

• Not checked yetcites [135]

In the paper

“Thus, post-processing approaches tend to apply transformations to model output to improve prediction fairness.”

In the paper

“Such that equal opportunity promotes that the TPR is the same across diferent groups.”

What the sources say

• Not checked yetcites [139]
• Not checked yetcites [239]

In the paper

“Similarly to equal opportunity, in addition to TPR equalized odds simultaneously considers FPR as well, i.e., the percentage of actual negatives that are predicted as positive.”

In the paper

“Using these properties, general fairness desiderata are categorized into three łnon-discriminationž criteria [16]: Independence aims for classiiers to make their scoring independent of the group membership:”

What the sources say

• Not checked yetcites [16]

In the paper

“Group-based fairness metrics essentially compare the outcome of the classiication algorithm for two or more groups.”

In the paper

“Over time, many diferent approaches have been suggested, most of which use metrics based on the binary classiication confusion matrix to deine fairness.”

In the paper

“While parity-based metrics typically consider variants of the predicted positive rate ( = 1), confusion matrixbased metrics consider additional aspects such as True Positive Rate (TPR), True Negative Rate (TNR), False Positive Rate (FPR), and False Negative Rate (FNR).”

In the paper

“Statistical/Demographic Parity: One of the earliest deinitions of fairness, this metric deines fairness as an equal probability of being classiied with the positive label [77,98,168,313].”

What the sources say

• Not checked yetcites [77]
• Not checked yetcites [98] source
• Not checked yetcites [168] source
• Not checked yetcites [313]

In the paper

“Most quantitative deinitions and measures of fairness are centered around three fundamental aspects of a (binary) classiier: 5 First, the sensitive variable that deines the groups for which we want to measure fairness.Second, the target variable .In binary classiication, this represents the two classes that we can predict: = 0 or = 1.”

In the paper

“In contrast to parity, it considers the ratio between unprivileged and privileged groups.”

In the paper

“[27] adjusts this concept by looking at relative accuracy rates across diferent groups.”

What the sources say

• Not checked yetcites [27]

In the paper

“A disadvantage of this notion, however, is that potential diferences between groups are not being taken into account.”

In the paper

“Its origins are in legal fairness considerations for selection procedures which sometimes use an 80% rule to deine if a process has disparate impact (ratio smaller than 0.8) or not [98].”

What the sources say

• Not checked yetcites [98] source

In the paper

“Test fairness/ calibration / matching conditional frequencies ( [69,139]): Essentially, test fairness or calibration wants to guarantee that the probability of = 1 is the same given a particular score.”

What the sources say

• Not checked yetcites [69] source
• Not checked yetcites [139]

In the paper

“1) many of the most contentious application areas that motivated the domain are binary decisions (hiring vs. not hiring; ofering a loan vs. not ofering a loan etc.);”

In the paper

“Treatment equality [27]: Treatment equality considers the ratio of False Negative Predictions (FNR) to False Positive Predictions.”

What the sources say

• Not checked yetcites [27]

In the paper

“Related to the Suiciency criterion, calibration-based metrics take the predicted probability, or score, into account, diferentiating them from metrics above that use predicted and actual values.”

In the paper

“[182] propose the concept of counterfactual fairness which builds on causal fairness models and is related to both individual and group fairness concepts.”

What the sources say

• Not checked yetcites [182]

In the paper

“We note an abundance of pre-and in-processing vs. post-processing methods and that many researchers increasingly develop intervention strategies that belong to multiple stages.”

In the paper

“2) quantifying fairness on a binary dependent variable is mathematically more convenient; addressing multi-class problems would add terms to the fairness quantity.”

In the paper

“As compared to group-based metrics which compare scores across diferent groups, individual and counterfactual fairness metrics do not focus on comparing two or more groups as deined by a sensitive variable, but consider the outcome for each participating individual.”

In the paper

“With a reliance on expressing fairness and bias mathematically, [76,128] argue that these deinitions often do not map to normative social, economic, or legal understandings of fairness.”

What the sources say

• Not checked yetcites [76]
• Not checked yetcites [128]

In the paper

“[270] note that many approaches to group fairness tackle only between-group issues, as a consequence worsening within-group fairness.”

What the sources say

• Not checked yetcites [270] source

In the paper

“In the absence of such a review, ofering advice on approach selection is unrealistic.”

In the paper

“Yet, determining which combination(s) of variables to blind is non-trivial and the risk of omitted variable bias should not be downplayed [71,160].”

What the sources say

• Not checked yetcites [71] source
• Not checked yetcites [160]

In the paper

“Both omission and immunity overlook relationships with proxy variables (as discussed in subsection 2.1, we note [291] as an exception here who omit proxies), which can result in increasing bias and discrimination [177], or indirectly concealing discrimination [88].”

What the sources say

• Not checked yetcites [177] source
• Not checked yetcites [88] source

In the paper

“They have also been criticized for not well examining how they would be applied in practice [253].”

What the sources say

• Not checked yetcites [253] source

In the paper

“Directed acyclic graphs (DAGs) are a common means to represent conditional independence assumptions between variables [155].”

What the sources say

• Not checked yetcites [155]

In the paper

“This means they can have high utility in studies of fairness in industry settings.”

In the paper

“Thus, causal methods are speciically well suited to identifying 1) proxies of sensitive variables as discussed in subsection 2.1; 2) which subgroups are most (un)fairly treated and diferentiate the types of bias exhibited [118]; and 3) provide transparency with respect to how (classiication) decisions were made [143].”

What the sources say

• Not checked yetcites [118] source

In the paper

“Initial results with data repair methods have shown to result in łdebiasedž classiiers which are robust to unseen test data, yet this process requires signiicant computational resources.”

What the sources say

• Not checked yetcites [251ś253]

In the paper

“[134] also learn proxy groupings.”

What the sources say

• Not checked yetcites [134]

In the paper

“[290] propose an iterative human-in-the-loop resampling approach where users can evaluate potential fairness issues of a model through interactive visualizations.”

What the sources say

• Not checked yetcites [290] source

In the paper

“Probabilistic veriication (e.g.[18]) of fairness measures via the sampling of sensitive variables has also been proposed to evaluate a trained model within some (small) conidence bound.”

What the sources say

• Not checked yetcites [18] source

In the paper

“[165], propose to preferentially sample (similar to oversampling) instances łclosež to a decision boundary (based on an initial model prototype to approximate a decision boundary) as these are most likely to be discriminated or favored due to underlying biases within the training data.”

What the sources say

• Not checked yetcites [165] source

In the paper

“Similar to other evaluation approaches (e.g.[10]), these approaches present fairness as a dichotomous outcome: a model is fair, or it isn't, as opposed to quantifying how (un)fair a model is.”

What the sources say

• Not checked yetcites [10] source

In the paper

“[148] extend this approach creating an ensemble of ensembles to operate on the protected groups.”

What the sources say

• Not checked yetcites [148]

In the paper

“Statistical hypothesis testing is employed to reveal whether models are signiicantly diferent with respect to fairness quantities or denote variable instability, i.e., when a model is not robustly fair when a given variable or set of variables are included.”

In the paper

“Thus, diferent candidate groupings are often evaluated via in-or post-processing methods to inhibit overitting, maximize some fairness metric(s), and/or prevent other theoretical violations.”

In the paper

“To illustrate transformation, [98] discuss transforming the distribution of SAT scores towards the median to łdegenderž the distribution into one which retains only the rank order for individuals, i.e, they remove information about protected variables from a set of covariates.”

What the sources say

• Not checked yetcites [98] source

In the paper

“Similarly, and especially for decoupled classiiers, is the challenge of selecting groups: some can be rarer than others [283] and as such a balance is needed to ensure groups are as atomic as possible but robust against gerrymandering [140,170].”

What the sources say

• Not checked yetcites [140]
• Not checked yetcites [170]

In the paper

“Common approaches in group formation are recursive partitioning (e.g.[170,298]) and clustering (e.g.: [70,148]) as (good) clusters well approximate underlying data distributions and subgroups.”

In the paper

“There are parallels between blinding (in the immunity sense) and independence mappings, as in many ways these two approaches share a common goal: creating independence from one or more speciic sensitive variables.”

In the paper

“[148] used clustering as a means to build stratiied samples of diferent subgroups within the data as an exercise in bagging for the training of fair ensembles.”

What the sources say

• Not checked yetcites [148]

In the paper

“Current transformation approaches operate mainly on numeric data, which is a signiicant limitation [98], yet approaches for other scenarios such as image classiications have been proposed [132].”

What the sources say

• Not checked yetcites [98] source

In the paper

“2) Understanding the relationship(s) between sensitive and potential proxy variables is hard [98], thus causal methods (subsection 4.2) may be useful precursor to transformation techniques;”

In the paper

“An alternative to retaining rank order is the use of distortion constraints (e.g.[88]) which seek to prevent mapping łhighž values to łlowž values and vice versa.”

What the sources say

• Not checked yetcites [88] source

In the paper

“[52,202] deine a similar approach yet model the transformation process with diferent assumptions and objectives.”

What the sources say

• Not checked yetcites [52] source
• Not checked yetcites [202]

In the paper

“[65,173,219] transform the output of a classiier in accordance to the identiication of unfair causal pathways, either by averaging [219], constraining the conditional distribution of the decision variable [173] or through counterfactual correction [65].”

What the sources say

• Not checked yetcites [65] source
• Not checked yetcites [173]
• Not checked yetcites [219] source

In the paper

“There are a number of challenges when applying transformation techniques: 1) the transformed data should not be signiicantly diferent from the original data, otherwise the extent of łrepairž can diminish the utility of the produced classiier [88,98,202] through data loss [125].”

What the sources say

• Not checked yetcites [88] source
• Not checked yetcites [98] source
• Not checked yetcites [202]
• Not checked yetcites [125]

In the paper

“Although largely a pre-processing method, transformation can also been applied within a post-processing phase.”

In the paper

“As an approach, this is similar to the idea of calibration (see subsection 4.10) and thresholding (see subsection 4.11) .”

In the paper

“It has been proposed as a mechanism to detect proxy variables, inluential variables [292] and counterfactual distributions [291].”

What the sources say

• Not checked yetcites [292]
• Not checked yetcites [291]

In the paper

“Often perturbation-based approaches are applied as a pre-processing step to prepare for an in-processing approach; often reweighing (e.g.[152]; see subsection 4.6), and/or regularisation / optimization (e.g.[283,291]; see subsection 4.7).”

What the sources say

• Not checked yetcites [152]
• Not checked yetcites [283]
• Not checked yetcites [291]

In the paper

“Many handle this by simply removing these samples, yet this may raise other methodological issues;”

In the paper

“It can also be applied to the test data upon the basis of strategies or probabilities learned on the training data.”

In the paper

“Referred to as data-massaging by [165,313], relabelling involves the modiication of the labels of training data instances so that the proportion of positive instances are equal across all protected groups.”

What the sources say

• Not checked yetcites [165] source
• Not checked yetcites [313]

In the paper

“Often, but not always, approaches seek to retain the overall class distribution.”

In the paper

“For example, [204] relabel the dependent variable (lip it from positive to negative or vice versa) if the data instance is determined as being discriminated against with respect to the observed outcome.”

What the sources say

• Not checked yetcites [204] source

In the paper

“6) transformation makes the model less interpretable [192,202], which may be at odds with data protection legislation;”

What the sources say

• Not checked yetcites [192] source
• Not checked yetcites [202]

In the paper

“Relabelling is also often used in counterfactual studies (e.g.[142,161,291]) which investigate if lipping the (in)dependent variable(s) afect the classiication outcome.”

What the sources say

• Not checked yetcites [142] source
• Not checked yetcites [161]
• Not checked yetcites [291]

In the paper

“[181] also note that some classiiers may be unafected by the presence or speciic nuances of some biases, and others may be negatively afected by altering the training data in an attempt to mitigate them.”

What the sources say

• Not checked yetcites [181] source

In the paper

“This is a relatively under-addressed area in the fairness literature, perhaps due to the introduction of SHAP valuesfoot_3 [203], which has seen an increase in attention from fairness researchers (e.g.[126]).”

In the paper

“[114,211] investigated visual mechanisms to better display łissuesž with data to users, yet these approaches generally lack support for novice users [24].”

What the sources say

• Not checked yetcites [114]
• Not checked yetcites [211]
• Not checked yetcites [24] source

In the paper

“This is a relatively under-addressed area in the fairness literature, perhaps due to the introduction of SHAP valuesfoot_3 [203], which has seen an increase in attention from fairness researchers (e.g.[126]).”

What the sources say

• Not checked yetcites [203]
• Not checked yetcites [126] source

In the paper

“As for transformation, modiication of the data via relabelling and perturbation is not always legally permissible [17] and changes to the data should be minimised [136,204].”

What the sources say

• Not checked yetcites [136] source
• Not checked yetcites [204] source

In the paper

“The application of sensitivity analysis in ML is often to measure the stability of a model [242].”

What the sources say

• Not checked yetcites [242] source

In the paper

“Yet sensitivity analysis (and explainable fairness in general) has been well motivated (although perhaps indirectly): [124] called for a better understanding of bias stemming from uncertainty, and [139] who stressed that assessment of data reliability is needed.”

What the sources say

• Not checked yetcites [124]
• Not checked yetcites [139]

In the paper

“Yet, relabelling and perturbation are not far from the principles of sensitivity analysis.”

In the paper

“Closely related to perturbation approaches is the use of sensitivity analysis (see [254]) to explore how various aspects of the feature vector afect a given outcome.”

What the sources say

• Not checked yetcites [254]

In the paper

“Whilst there are a number of papers that harness perturbation (it is not always referred to as perturbation) in the ML literature, this approach appears more prevalent in the discrimination-aware data mining literature where it is often used as a means of privacy preservation.”

In the paper

“[80] proposed the perturbation of feature vectors to measure the efect on model performance of speciic interventions.”

What the sources say

• Not checked yetcites [80]

In the paper

“Weights can be introduced for multiple purposes: 1) to indicate a frequency count for an instance type (e.g.[52]), 2) to place lower/higher importance on łsensitivež training samples (e.g.[53,152,165]), or 3) to improve classiier stability (e.g.[181]).”

What the sources say

• Not checked yetcites [181] source

In the paper

“Weights can be introduced for multiple purposes: 1) to indicate a frequency count for an instance type (e.g.[52]), 2) to place lower/higher importance on łsensitivež training samples (e.g.[53,152,165]), or 3) to improve classiier stability (e.g.[181]).”

What the sources say

• Not checked yetcites [52] source

In the paper

“For example, [165] seek to assign weights that take into consideration the likelihood of an instance with a speciic class and sensitive value pairing (pre-processing).”

What the sources say

• Not checked yetcites [165] source

In the paper

“Whereas, [181] irst build an unweighted classiier, learn the weights of samples, then retrain their classiier using these weights (in-processing).”

What the sources say

• Not checked yetcites [181] source

In the paper

“A similar approach is taken by [152] who identify sensitive training instances (pre-processing), but then learn weights for these instances (in-processing) to optimize for the chosen fairness metric.”

What the sources say

• Not checked yetcites [152]

In the paper

“Some notable exceptions emphasise: 1) empirical risk subject to fairness constraints or welfare conditions (e.g.[87,142]), 2) TPR/FPR of protected groups (e.g.[19]), 3) stability of fairness (e.g.[145]), or 4) counterfactual terms (e.g.[85]).”

What the sources say

• Not checked yetcites [19]

In the paper

“Applied to fairness, regularization means adding penalty terms to penalize the classiier for discriminatory practices [168].”

What the sources say

• Not checked yetcites [168] source

In the paper

“Thus, it is not hypothesis (or model) driven, but data driven [19] and based upon the notion(s) of fairness considered.”

What the sources say

• Not checked yetcites [19]

In the paper

“Some notable exceptions emphasise: 1) empirical risk subject to fairness constraints or welfare conditions (e.g.[87,142]), 2) TPR/FPR of protected groups (e.g.[19]), 3) stability of fairness (e.g.[145]), or 4) counterfactual terms (e.g.[85]).”

What the sources say

• Not checked yetcites [145]

In the paper

“Some notable exceptions emphasise: 1) empirical risk subject to fairness constraints or welfare conditions (e.g.[87,142]), 2) TPR/FPR of protected groups (e.g.[19]), 3) stability of fairness (e.g.[145]), or 4) counterfactual terms (e.g.[85]).”

What the sources say

• Not checked yetcites [85]

In the paper

“[218] also approached this via reinforcement learning.”

What the sources say

• Not checked yetcites [218]

In the paper

“Yet, these approaches can also include other constraints, and/or reduce the problem to a cost-sensitive classiication problem (e.g.[3,120,221]).”

What the sources say

• Not checked yetcites [3]
• Not checked yetcites [120]
• Not checked yetcites [221]

In the paper

“3) diferent regularization terms and penalties have diverse results on diferent data sets, i.e. this choice can have qualitative efects on the trade-of between accuracy and fairness [26].”

What the sources say

• Not checked yetcites [26]

In the paper

“[79,120,221] summarizes a number of additional constraint types as precision or budget constraints to address the accuracy-fairness trade-of (often expressed as utility or risk functions e.g.[77,135]); quantiication or coverage constraints to capture disparities in class or population frequencies; churn constraints capturing online learning scenarios and enforcing that classiiers do not difer sign”

What the sources say

• Not checked yetcites [79]
• Not checked yetcites [120]
• Not checked yetcites [221]
• Not checked yetcites [109] source
• Not checked yetcites [145]

In the paper

“This idea of łrobustness of fairnessž is often not considered by many works and would warrant better inclusion in future fairness studies.”

In the paper

“With the goal of better fairness generalization, the approach can also be used to quantify the robustness of fairness.”

What the sources say

• Not checked yetcites [103]

In the paper

“In light of this, [103] suggest a robustness framework for fairness that ensures fairness over diferent training samples.”

What the sources say

• Not checked yetcites [103]

In the paper

“Aside from this, [117,313] note that regularization as a mechanism is fairly generic, and can lead to a lack of model robustness and generalizability.”

What the sources say

• Not checked yetcites [117] source
• Not checked yetcites [313]

In the paper

“[145] argue that stability of fairness can be addressed through regularization and present corresponding empirical evidence through extensions of [168,307].”

What the sources say

• Not checked yetcites [145]
• Not checked yetcites [168] source
• Not checked yetcites [307] source

In the paper

“[145] argue that stability of fairness can be addressed through regularization and present corresponding empirical evidence through extensions of [168,307].”

What the sources say

• Not checked yetcites [145]

In the paper

“Key challenges for regularization approaches are: 1) they are often non-convex in nature or achieve convexity at the cost of probabilistic interpretation [119];”

What the sources say

• Not checked yetcites [119] source

In the paper

“When applied to applications of fairness in ML, an adversary instead seeks to determine whether the training process is fair, and when not, feedback from the adversary is used to improve the model [58].”

What the sources say

• Not checked yetcites [58]

In the paper

“Additionally, in a Deep Neural Network setting, [115] show that diferent trade-ofs between accuracy and fairness can be achieved by selective dropout of neurons.”

What the sources say

• Not checked yetcites [115]

In the paper

“An advantage of adversarial approaches is that they can consider multiple fairness constraints [289], often treating the model as a blackbox [206].”

What the sources say

• Not checked yetcites [289]
• Not checked yetcites [206]

In the paper

“This is often formulated as a multi-constraint optimization problem considering many of the constraints as discussed in subsection 4.7.”

In the paper

“Work has also started applying the notions of causal and counterfactual fairness to adversarial learning (e.g.[299]).”

What the sources say

• Not checked yetcites [299]

In the paper

“As a reinforcement learning framework, bandits are motivated on the need for decisions to be made in an online manner [156] (although not always, e.g.[215]), and that decision makers may not be able to deine what it means to be łfairž but that they may recognize łunfairnessž when they see it [116].”

What the sources say

• Not checked yetcites [116]

In the paper

“This allows researchers to frame the fairness problem as a stochastic multi-armed bandit, assigning either individuals or groups of łsimilarž individuals to arms, and represent fairness as regret [156,201].”

What the sources say

• Not checked yetcites [156]
• Not checked yetcites [201]

In the paper

“As a reinforcement learning framework, bandits are motivated on the need for decisions to be made in an online manner [156] (although not always, e.g.[215]), and that decision makers may not be able to deine what it means to be łfairž but that they may recognize łunfairnessž when they see it [116].”

What the sources say

• Not checked yetcites [156]

In the paper

“The use of generative adversarial networks (GAN) with fairness considerations also permit applications within unstructured (for example multimedia) data or more generally as a generative process of creating an łunbiasedž dataset using a number of samples.”

In the paper

“It is important to highlight that the diference between the application of bandits to fair classiication and fair recommendation (subsection 5.2) is often hard to distinguish.”

In the paper

“The application of bandits to ML fairness has proliferated recently.”

In the paper

“The main notions of fairness applied in the application of bandits are meritocratic fairness [127,156,157] (group agnostic), subjective fairness [201] (emphasises fairness in each time period ) and counterfactual fairness (e.g.[146]).”

What the sources say

• Not checked yetcites [127]
• Not checked yetcites [156]
• Not checked yetcites [157] source
• Not checked yetcites [201]
• Not checked yetcites [146]

In the paper

“The literature has proposed various approaches to handle the impasse of achieving calibration and other fairness measures.”

In the paper

“Calibrating a ML model for multiple protected groups and/or using multiple fairness criteria at once has been shown to be impossible [69,140,174,178,200,201,223,239].”

What the sources say

• Not checked yetcites [69] source
• Not checked yetcites [140]
• Not checked yetcites [174]
• Not checked yetcites [178]
• Not checked yetcites [200]
• Not checked yetcites [201]
• Not checked yetcites [223]
• Not checked yetcites [239]

In the paper

“One approach has been to apply a randomization post-processing process to try and achieve a balance between accuracy and fairness, yet [76,139,178,223,239] discuss a number of shortcomings of this approach.”

In the paper

“As such, a calibrated model does not inhibit biases of decision makers, but ensures that risk estimates for various (protected) subgroups carry the same meaning [239].”

What the sources say

• Not checked yetcites [239]

In the paper

“[239] even note that the goals of low error and calibration are competing objectives for a model.”

What the sources say

• Not checked yetcites [239]

In the paper

“[223] also note that this approach is Pareto sub-optimal, and instead propose a cost-based approach to balance calibration and error parity.”

What the sources say

• Not checked yetcites [223]

In the paper

“Notably, the individuals who are randomized are not necessarily positively impacted, and the overall accuracy of the model can be adversely afected.”

What the sources say

• Not checked yetcites [76]
• Not checked yetcites [139]
• Not checked yetcites [178]
• Not checked yetcites [223]
• Not checked yetcites [239]

In the paper

“[170,175] undertake a similar approach under diferent settings.”

What the sources say

• Not checked yetcites [170]
• Not checked yetcites [175] source

In the paper

“[140] instead seek to achieve approximate calibration (i.e. to guarantee calibration with high probability) using a multi-calibration approach that operates on identiiable subgroups to balance individual and group fairness measures even for small samples: a speciic challenge for achieving calibration [200].”

What the sources say

• Not checked yetcites [140]

In the paper

“[223] also note that this approach is Pareto sub-optimal, and instead propose a cost-based approach to balance calibration and error parity.”

What the sources say

• Not checked yetcites [223]

In the paper

“[201] have proposed a bandit-based approach to calibration.”

What the sources say

• Not checked yetcites [201]

In the paper

“[200] suggest that calibration is suicient as a fairness criterion if the model is unconstrained.”

What the sources say

• Not checked yetcites [200]

In the paper

“Computing by hand can introduce new biases without training [286], as fairness is typically not monotonic, thus assigning a threshold value may be quite arbitrary [270].”

What the sources say

• Not checked yetcites [286] source
• Not checked yetcites [270] source

In the paper

“The underlying idea here is to incentivize good performance (in terms of both fairness and accuracy) across all classes and groups.”

In the paper

“Such instances are considered to be ambiguous, and therefore potentially inluenced by bias [167].”

What the sources say

• Not checked yetcites [167] source

In the paper

“Threshold values are often computed with respect to accuracy, but cases of class imbalance would invalidate this [148].”

What the sources say

• Not checked yetcites [148]

In the paper

“note that there is a suicient need for fairness researchers to also focus on other ML problems.”

In the paper

“An aspect that distinguishes fairness considerations in recommender systems is that fairness can be seen as multi-sided concept that can be relevant for both users (who consume the recommendations) and items.”

In the paper

“The motivation for using transfer learning is typically in response to an observable covariate shift between the source (training) and target distributions.”

What the sources say

• Not checked yetcites [31]
• Not checked yetcites [240]
• Not checked yetcites [273]

In the paper

“However, when implementing fairness measures, we must emphasize either fairness or model performance as improving one can often detriment the other [27,54,77,89,135,139,319].”

What the sources say

• Not checked yetcites [27]
• Not checked yetcites [54]
• Not checked yetcites [77]
• Not checked yetcites [89] source
• Not checked yetcites [135]
• Not checked yetcites [139]
• Not checked yetcites [319]

In the paper

“ML practitioners can measure performance in a multitude of ways, and there has been much discussion concerning the choice of diferent performance measures and approaches [84,267].”

What the sources say

• Not checked yetcites [84]
• Not checked yetcites [267] source

In the paper

“Yet, there is a cost here, unlike other approaches to fairness in ML, fair clustering has signiicant computational costs associated to it.”

In the paper

“As noted by [98], however, a reduction in accuracy may in fact be the desired result, if it was discrimination that raised accuracy in the irst place.”

What the sources say

• Not checked yetcites [98] source

In the paper

“Approaches that utilize transfer learning do so in combination with other methods.”

In the paper

“E.g., [237] notes that women have difering views in the inclusion / exclusion of gender as a protected variable to men.”

What the sources say

• Not checked yetcites [237]

In the paper

“The choice of performance measure(s) itself may even harbor, disguise, or create new underlying ethical concerns.”

In the paper

“Thus, improving fairness in ML is challenging and simultaneously there are many diferent notions for researchers and practitioners to navigate.”

In the paper

“[270] also note that often approaches to group fairness tackle between-group issues, worsening within-group issues through this choice.”

What the sources say

• Not checked yetcites [270] source

In the paper

“This is corroborated by [266] who note an over-emphasis in the literature on speciic measures of fairness and insuicient dialogue between researchers and afected communities.”

What the sources say

• Not checked yetcites [266]