Bias Mitigation for Machine Learning Classifiers: A Comprehensive Survey

1 INTRODUCTION

• ML algorithms can replicate or amplify real-world data bias, causing unequal treatment; the survey treats fairness as the absence of bias.Biased behavior may also risk violating anti-discrimination laws.Source: §1 · pp. 1–2
• Bias mitigation methods intervene through pre-processing of training data, in-processing during model training, or post-processing of trained models.Source: §1 · p. 2
• Practitioners lack clear guidance on evaluating new mitigation mechanisms, including which measures, datasets, and comparison methods to use, as Pessach and Shmueli [287] note.Source: §1 · p. 2
• The survey analyzes 341 publications on bias mitigation for ML classifiers, covering methods, evaluation datasets, fairness metrics, benchmarking, and future research challenges.The authors also make the collected paper repository public to support replication and manual investigation.Source: §1 · pp. 2–3

2 RELATED SURVEYS

• Surveys by Mehrabi et al. and Pessach and Shmueli [287] covered bias, fairness definitions and metrics, mitigation methods, and datasets; Dunkelau and Leuschel classified methods by intervention stage.Caton and Haas [52] conducted the most exhaustive categorization of bias mitigation methods to date and also presented fairness metrics and platforms.Source: §2 · p. 3
• Fairness-definition and metric surveys include collections for classification, indirect-discrimination measures [417], and historical parallels between past and current fairness research.Hutchinson and Mitchell connected test fairness in education and hiring in the 1960s and 1970s to current individual- and group-fairness approaches and confusion-matrix measures.Source: §2 · p. 3
• Le Quy et al. surveyed 15 frequently used fairness datasets, describing their features and relationships with sensitive attributes.Pessach and Shmueli also listed datasets and metrics used by 27 bias mitigation methods.Source: §2 · p. 3
• Other fairness surveys address sequential decision algorithms [407], criminal justice [38], graph representations [406], ML testing, software engineering [68] [336], and natural language processing [44] [338].Source: §2 · p. 3
• Previous surveys mentioned bias mitigation but did not systematically cover its evaluation, such as benchmarking and dataset use.Dunkelau and Leuschel focused on implementation and algorithms; Pessach and Shmueli listed datasets and metrics for 27 methods but offered no actionable insights for developers.Source: §2 · p. 3

3.1 Search Methodology

• The search combines preliminary collection, six-repository keyword searches, eligibility screening, and backward snowballing to identify relevant work.The survey starts from 100 unique publications found in four existing surveys; eligible work describes human biases, addresses classification, and uses tabular data, with title, abstract, and full-body checks.Source: §3.1 · p. 4

3.2 Selected Publications

• The survey gathered 341 publications through its search procedure, adding 32 after corresponding authors reviewed descriptions of their work.The second repository search, conducted in July 2022, targeted 2022 publications; Google Scholar searches were limited to the first 1,000 results each.Source: §3.2 · p. 5
• 2018 saw publications more than double from 20 to 46, and NeurIPS, ICML, and AAAI had the most publications among named venues.Publications increased over the years; each year from 2009–2016 had fewer than 10. FAccT had 13 publications and AIES 12.Source: §3.2 · pp. 5–6, Figure 2 · p. 6

3.3 Visibility

• AI, Data, and other venues had the most publications (139, 59, and 68, respectively), while Fairness venues achieved the highest average citations in 2018, 2019, and 2021.Among 19 works cited more than 500 times, only two were outside AI or Data venues: Dwork et al. [109] and Zhang et al..Source: Figure 2 · p. 6, §3.3 · p. 7, Figure 3 · p. 7

3.4 Limitations

• The survey focuses on algorithmic fairness in ML models and bias mitigation for tabular data, excluding methods applied solely to visual or textual tasks.Its overview reflects research-proposed methods, often evaluated on public data, and results may not transfer to real-world applications [32] [99].Source: §3.4 · p. 8

4 ALGORITHMS

• The survey classifies methods by intervention stage and organizes them into 13 bias-mitigation categories based on existing taxonomies.The authors follow taxonomies devised by Dunkelau and Leuschel and Caton and Haas [52].Source: §4 · p. 8
• Of 341 publications, 123 used pre-processing, 212 in-processing, and 56 post-processing methods; 70 publications examined more than one type.A single publication could count in up to three types, such as when pre-processing was combined with changes to the training procedure.Source: §4 · p. 8

4.1 Pre-processing Bias Mitigation Methods

• Pre-processing combats bias by changing training data, including ground-truth labels or other feature values.The survey lists 123 publications using pre-processing methods.Source: §4.1 · p. 8, §4.1.1 · p. 8

4.1.1 Relabelling and Perturbation.

• Relabelling methods include massaging, which switches labels for rank-selected instances near the decision boundary, while perturbation adjusts feature values across groups.Perturbation can preserve within-group ranks [118], modify multiple variables using conditional models [240], or iteratively transform the most bias-prone attribute until bias falls below a threshold.Source: §4.1.1 · pp. 8, 10

4.1.2 Sampling.

• Sampling methods alter sample distributions or instance impact through downsampling, oversampling, SMOTE-generated minority-group instances, preferential sampling, and reweighing.Reweighing assigns weights by label and protected attribute; higher weights increase losses when instances are misclassified, and weighted instances may be resampled or used directly when supported.Source: §4.1.2 · p. 10
• Other sampling approaches generate discrimination-free training data, modify causal networks to create new datasets, or augment data by swapping protected attributes.Causal-network-based data generation is described by Zhang et al., while Sharma et al. duplicate existing datasets and swap each instance’s protected attribute.Source: §4.1.2 · p. 10

4.1.3 Latent variables.

• Latent variables can represent labels or group membership, and proxy group variables with fairness constraints can address datasets missing group labels.Calders and Verwer used clustering and expectation maximization to identify instances for positive or negative latent labels; Gupta et al. [140] created a latent proxy for group membership.Source: §4.1.3 · pp. 10–11

4.1.4 Representation.

• Representation learning seeks to reduce bias while preserving information, including LFR, optimization-based methods, and removal or addition of features.LFR optimizes to remove protected-attribute information while minimizing loss of non-sensitive information; feature methods include removing discriminating features and creating nonlinear transformations while dropping biased features.Source: §4.1.4 · p. 11

4.2 In-processing Bias Mitigation Methods

• The survey identifies 212 publications applying in-processing bias mitigation during training.Tables 4 and 5 provide further publication details; the survey also refers readers to Wan et al. for 38 in-processing approaches across ML tasks.Source: §4.2 · p. 11

4.2.1 Regularization and Constraints.

• Regularization penalizes discrimination in the loss, while constraints set bias levels that training must not breach.For decision trees, Kamiran et al. considered fairness alongside accuracy in leaf splitting, using non-discriminatory splits or criteria combining accuracy and discrimination changes; Celis et al. [53] proposed a meta-algorithm that takes a fairness constraint as input. These methods also cover individual fairness, multiple sensitive attributes, and additional regularization such as alongside the L2 norm.Source: §4.2.1 · pp. 11, 14

4.2.2 Adversarial Learning.

• Adversarial learning trains a classifier and an adversary to improve their competing objectives, including predicting labels while limiting sensitive-attribute prediction.Zhang et al. used Logistic Regression predictor and adversary models with Demographic Parity, Equality of Odds, and Equality of Opportunity; Beutel et al. [40] used a neural network with shared layers to predict labels and sensitive attributes. Lahoti et al. proposed adversarially reweighted learning, while adversarial training has also been used for poisoning robustness, individual fairness, and training-data reweighing.Source: §4.2.2 · p. 14

4.2.3 Compositional.

• Compositional approaches train multiple classification models, then predict using a specific model for each population group [46] [340] or by voting across models in an ensemble [258].Decoupling reduces training data per classifier; Dwork et al. trained on respective-group and other-group data, with the other groups’ data given reduced weight. Training models with different fairness goals creates a Pareto front of solutions, allowing practitioners to choose a fairness-accuracy trade-off [43] [258].Source: §4.2.3 · p. 14

4.2.4 Adjusted Learning.

• Adjusted learning changes training procedures or algorithms, with methods addressing fairness through active learning, deferral, group risk, calibration, and privacy.Examples include active-learning frameworks that gather information or select instances for labeling, rejection learning that defers to external decision makers, APStar’s minimization of maximum group risk, multicalibration’s iterative correction of the most biased subgroup, distributionally robust optimization of worst-case group risk, and privacy-aware Logistic Regression training.Source: §4.2.4 · p. 15

4.3 Post-processing Bias Mitigation Methods

• 56 publications applied post-processing methods, making them the least frequently used intervention type covered in the survey.These methods are applied after a classification model has been successfully trained.Source: §4.3 · p. 15

4.3.1 Input Correction.

• Only two publications corrected test inputs, and both used perturbations; Li et al. also applied the same procedure during pre-processing.Adler et al. used perturbation only at the post-processing stage.Source: §4.3.1 · p. 15

4.3.2 Classifier Correction.

• Hardt et al. [144] optimized classifier modifications for Equalized Odds and Equality of Opportunity, but Woodworth et al. warned they can reduce accuracy when the loss is not strictly convex.Woodworth et al. also adapted Hardt et al.’s approach alongside training constraints.Source: §4.3.2 · p. 16
• Pleiss et al. split classifiers by group and probabilistically replaced unprivileged-group predictions with the class mean to balance error rates; Noriega-Campero et al. followed this calibration approach.Source: §4.3.2 · p. 16
• Decision-tree corrections relabel leaves greedily by fairness gain per accuracy loss or optimize branching thresholds with a mixed integer program.Source: §4.3.2 · p. 16
• Multiaccuracy Boost iteratively adapts a black-box classifier according to its predictive performance, using a learning algorithm in a boosting-like process.Source: §4.3.2 · p. 17

4.3.3 Output Correction.

• Post-processing methods mitigate biased outputs by adjusting rule confidence, thresholding predictions, or relabeling likely biased cases near decision boundaries or with individual-bias detectors; group-dependent thresholds are another option [15] [78] [284] [391] [392].Source: §4.3.3 · p. 17
• For causal models, Chiappa [72] corrected predictions to match a counterfactual population group, removing sensitive-attribute effects while preserving other information.Source: §4.3.3 · p. 17

Combined Approaches

• Overall, 70 publications applied multiple techniques at the same time, such as two pre-processing methods or methods from different stages.The survey contrasts these with the majority of publications, which proposed a single type of bias mitigation method.Source: §4.4 · p. 17
• Among 70 publications using multiple techniques, 60 (86%) applied in-processing, 38 (54%) pre-processing, and 22 (31%) post-processing methods.These categories can overlap because a publication may use methods at multiple stages.Source: §4.4 · p. 17
• Among 26 publications combining methods at the same stage, 19 used multiple in-processing methods, 17 of which included regularization or constraints.Seven used multiple pre-processing methods; five of those combined sampling and relabeling.Source: §4.4 · p. 17
• 47 publications combined methods across different development stages, and three addressed all three stages [140].The survey notes that bias mitigation methods can be used in conjunction [129].Source: §4.4 · p. 17
• Examples of all-stage approaches include Calders and Verwer, Gupta et al. [140], and Iosifidis et al..Calders and Verwer add a latent variable, train models for each sensitive-attribute value, and adjust output probabilities; Gupta et al. use a proxy group variable and fairness constraints before debiasing an existing classifier. Iosifidis et al. use an AdaBoost ensemble trained on balanced group-label samples and modify predictions with group-dependent thresholds.Source: §4.4 · pp. 17–18

• Logistic Regression appears in 140 unique publications, Neural Networks in 102 (mostly in-processing), and Linear Regression in 22.Source: §4.5 · p. 19
• Decision Trees appear in 36 publications and Random Forests in 45; Decision Tree variants include Hoeffding trees, C4.5, J48, and Bayesian random forests.Source: §4.5 · p. 19
• 70% of publications applied their bias mitigation method to only one classification model.Source: §4.5 · p. 19, Figure 4 · p. 20
• Only 24% of publications with in-processing methods consider more than one classification model, compared with 35% for pre-processing and 43% for post-processing.In-processing methods are model-specific and directly modify the training procedure, whereas pre-processing and most post-processing methods can be developed independently of the classification model and applied to multiple models.Source: §4.5 · p. 20

5.3 Synthetic Data

• The overview reports 56 publications creating synthetic or semi-synthetic data; the detailed breakdown reports 54 synthetic-data publications and 3 using semi-synthetic data.Synthetic datasets range from hundreds to 100,000 or more data points, and some publications do not report dataset size [96] [29] [75] [248]. One commonly adapted approach generates 4,000 binary labels, adds two Gaussian-distributed user features, and draws the sensitive attribute from a Bernoulli distribution.Source: §5 · p. 20, §5.3 · p. 22

5.1 Dataset Usage

• Of 81 non-synthetic datasets, 73% were used by only one or two publications.The remaining 22 datasets were each used in at least three publications.Source: §5.1 · pp. 20–21
• The Adult dataset was used by 77% of empirical publications, making it the most frequently used dataset.Table 9 covers datasets used in at least three publications. The authors recommend Adult and COMPAS for comparability but caution that threshold choices affect Adult’s external validity [99] and that COMPAS should be used with domain context [32].Source: Table 9 · p. 21, §5.1 · p. 21

5.2 Dataset Frequency

• Two datasets was the most common evaluation choice, and 90% of publications used four or fewer.The authors recommend evaluating on at least three datasets, preferably diverse in domain, dimensionality, or protected attributes.Source: §5.2 · p. 22, §2.7 · p. 22

5.4 Data-split

• Among 324 empirical publications, 232 reported data-split information and 143 reported the number of runs.Default dataset splits may explain some missing split-size information; Adult, for example, has a predefined 70%-30% train-test split.Source: §5.4 · p. 23
• For non-k-fold splits, 80%-20% (39 times) and 70%-30% (35 times) were most common; 47 publications used k-fold cross-validation.Experiments were repeated 10 times in 54 of 143 cases and 5 times in 42; 36 publications used validation splits, most commonly 60% training, 20% testing, and 20% validation. Sixteen publications used different train-test splits across datasets, while streaming, federated, multi-source, and sequential methods may require different split approaches [7] [114] [152] [284] [292] [158] [20] [410].Source: §5.4 · p. 23

6.1 Definitions Based on Labels in Dataset

• Mean Difference was the most popular dataset-label metric, used in six publications; for elift and slift, perfect fairness corresponds to ε = 0.These metrics measure disparities in positive labels across population groups.Source: §6.1 · p. 25

6.2 Definitions Based on Predicted Outcome

• Demographic Parity compares groups’ proportions of positive predictions using a difference or ratio; Žliobaite [417] advised against ratios as harder to interpret.The p-rule compares the two possible ratios and selects the minimum; absolute differences can be used when the direction of bias is unimportant [94].Source: §6.2 · p. 25

6.3 Definitions Based on Predicted and Actual Outcomes

• Equality of Opportunity was the most-used predicted-and-actual-outcome metric (90 uses), followed by Equalized Odds (52); the former compares groups’ TPRs, while the latter compares TPRs and FPRs.Accuracy differences were used 29 times, and FPR and TNR were used 38 and 8 times, respectively.Source: §6.3 · p. 26

6.4 Definitions Based on Predicted Probabilities and Actual Outcome

• Probability-based metrics include Strong Demographic Disparity and Strong Pairwise Demographic Parity, which compute parity across thresholds, and fair calibration, which returns a binary fairness result.Fair calibration bins each subgroup’s predicted probabilities into 10 equally sized groups and checks calibration within and across groups.Source: §6.4 · pp. 26–27

6.5 Definitions Based on Similarity

• Similarity-based fairness seeks equal predictions for similar individuals; consistency, the most-used similarity metric, appeared in four publications.Consistency compares an individual’s prediction with those of their k-nearest neighbors in input space.Source: §6.5 · p. 27

6.6 Causal Reasoning

• Causal metrics use causal graphs to assess sensitive attributes’ relationships with outcomes; counterfactual fairness appeared in four publications, and sensitive information was predictable from representations in nine.Direct discrimination depends on sensitive attributes, while indirect discrimination can arise through statistical dependencies between protected and non-protected attributes; the Prejudice Index and Normalized Prejudice Index measured indirect discrimination four times.Source: §6.6 · pp. 27–28

7 BENCHMARKING

• 308 of 324 experimental publications performed benchmarking; the remaining 15 did not compare results against any benchmark.The survey’s full corpus comprised 341 publications.Source: §7 · p. 28

7.1 Baseline

• The Original Model was the most-used baseline, appearing in 254 of 308 experiments.Other baselines were suppressing (30 experiments) and random baselines (13); four publications used a constant classifier [260], which gives each group the same treatment. Suppression alone may leave unfairness because other features can correlate with the protected attribute; Kamiran et al. also suppressed the k most-correlated features.Source: §7.1 · pp. 28–29

7.2 Benchmarking Against Bias Mitigation Methods

• 44% of pre-processing methods (50 of 114) did not benchmark against other mitigation methods, versus 36% of in-processing methods (66 of 184) and 31% of post-processing methods (16 of 52).Each method type was more likely to benchmark against methods of the same type; 137 mitigation methods were used as benchmarks, and the five most frequently used included all three types.Source: §7.2 · p. 29, Figure 6 · p. 30

7.3 Benchmarking Against Fairness-Unaware Methods

• Overall, 51 publications benchmarked against fairness-unaware methods; 23 were pre-processing and 27 in-processing, while six post-processing publications were reported, with at least one per approach type.Seven approach types benchmarked against fairness-unaware methods; this was rare among post-processing methods.Source: §7.3 · pp. 29–30

7.4 Source Code Availability

• 192 of 341 publications (56%) shared source code, and the proportion reached 71% in 2020.The proportion was below 50% in 2018 (46%) and 2019 (49%). Themis-ML [31], Fairlearn, and AIF 360 [36] implement three, four, and 13 methods, respectively; the survey also provides available dataset and implementation links in its repository.Source: §7.4 · pp. 30–31, Figure 7 · p. 31

8.1.1 Fairness Definitions.

• The literature uses 109 fairness metrics, making a common set difficult to establish; evaluation also needs to address multi-class problems and non-binary sensitive attributes, studied by only 15 publications.Metrics should represent the problem at hand; fairness work must also consider trade-offs among accuracy and multiple fairness metrics, and whether explainable discrimination is acceptable when fairness criteria conflict.Source: §8.1.1 · p. 32
• Formal fairness guarantees can bound bias for a specified metric and help when test sets may not represent real-world or unknown data distributions; eight publications identify guarantees as future work.The passage also notes that interpretability and explainability can help, particularly in domains subject to legislation and regulatory controls.Source: §8.1.2 · p. 32

8.1.2 Fairness Guarantees.

• Protected attributes are unavailable in practice, making their reliability and access a pressing concern raised by nine publications.The passage identifies access to protected attributes as a dataset-related challenge.Source: §8.1.3 · p. 32

8.1.3 Datasets.

• Dataset challenges include potentially biased annotations, limited protected-attribute diversity, and uncertainty about whether results on public datasets transfer as data distributions change.Possible responses include unbiased collection, label debiasing, expert annotation, and collecting more data, though subgroup-targeted collection may itself be biased; distribution shifts may require continuous mitigation.Source: §8.1.3 · pp. 32–33, §8.1.4 · p. 33

8.1.4 Real-world Applications.

• Real-world mitigation can miss relevant subgroups or apply inappropriate metrics, risking further unfairness or fairwashing; user studies can assess whether practitioners will use proposed methods.Datasets may simplify protected attributes through binarization, and low bias scores alone do not establish fairness; developers and policymakers could be asked about comfort with training-data augmentation.Source: §8.1.4 · p. 33

8.1.5 Extension of Experiments.

• Experiments could expand beyond current practice: 40 publications propose further metrics, 12 additional classification models, and only eight further datasets.The survey reports 81 datasets used at least once and an average of 2.7 datasets per publication; proposed extensions also include recommendation, ranking, and clustering tasks.Source: §8.1.5 · p. 33

8.2 Research Opportunities

• In-processing methods have almost twice as many publications as pre-processing and nearly four times as many as post-processing; input modification after training appears in only two publications.Pre- and post-processing methods are classifier-agnostic, yet Bandits have been studied only with in-processing methods. Combining pre- and post-processing has appeared four times, while no publication applied multiple post-processing methods.Source: §8.2 · p. 34

8.2 Research Opportunities.

• Evaluation opportunities include using multiple metrics across six categories, reporting accuracy for fairness–accuracy trade-offs, and benchmarking against baselines and relevant prior methods.The survey describes 109 metrics and reports source-code implementations in 192 studies that could serve as benchmarks; it also notes evaluations can draw on up to 81 datasets.Source: §8.2 · p. 34

9 CURRENT BEST PRACTICES / RECOMMENDATIONS

• Check existing approaches to confirm that the bias mitigation method under evaluation is novel.The authors also note that applying methods to more datasets or using more evaluation metrics can support result validity.Source: §9 · p. 34
• Apply the method to at least three diverse datasets, consider criticism when claiming real-world impact, and state each dataset’s protected attributes.The recommendation is to account for diversity and criticism when making claims about real-world impact.Source: §9 · p. 35
• Evaluate the method with at least two fairness metrics and a performance metric, using different metric types to reduce correlation among fairness metrics.Accuracy is given as an example of a performance metric.Source: §9 · p. 35
• Benchmark against the original model and similar existing mitigation methods, and test multiple classifiers, especially for pre- or post-processing methods.Logistic regression and neural networks are frequently used classifiers.Source: §9 · p. 35
• Repeat standard-split experiments at least 10 times, and share code and numerical results, especially when presenting results in bar charts.The passage gives 70% splits and predefined data splits as examples of standard training splits.Source: §9 · p. 35

10 CONCLUSION

• In-processing was most common among 123 pre-processing, 212 in-processing, and 56 post-processing methods; loss-function changes led, with 74 constraint-based and 51 regularization approaches.The survey devised 13 categories across the three method types; post-processing input correction appeared only twice.Source: §10 · p. 35
• 81 datasets appeared across the publications, but evaluations used an average of only 2.7; Adult was most popular, used by 77% of publications.Source: §10 · p. 35
• 109 unique fairness metrics were identified across six categories, with the most frequently used belonging to predicted-outcome and predicted-and-actual-outcome definitions.Source: §10 · p. 35
• 82% of methods used the original model as a baseline, 51 publications benchmarked against fairness-unaware methods, and 192 of 341 publications made source code available.The survey also found three frameworks implementing and making existing bias mitigation methods available [31] [36].Source: §10 · p. 35
• The survey identifies unresolved challenges: synthesizing fairness metrics amid a lack of consensus, accounting for fairness guarantees, and identifying relevant population groups and reliable datasets in practice.It aims to help researchers and practitioners understand existing approaches and support development of new methods.Source: §10 · pp. 35–36