Bias Mitigation in ML Models

Updated 30 November 2025

Bias Mitigation Techniques are algorithmic, statistical, and procedural methods designed to reduce disparities in predictive model performance across diverse groups.
They are categorized into pre-processing, in-processing, and post-processing methods, each with unique trade-offs and practical implications.
Implementing these techniques successfully requires rigorous tuning, multi-metric evaluation, and active stakeholder engagement to address fairness challenges.

Bias mitigation techniques are algorithmic, statistical, and procedural interventions designed to reduce or eliminate disparate predictive performance or allocation across socially salient groups in ML and AI models. These techniques address biases arising from spurious correlations, sample imbalances, labeling artifacts, or structural asymmetries in feature distributions. The bias mitigation literature has evolved rapidly, resulting in a diverse toolset including pre-processing, in-processing, and post-processing methods, each with distinct theoretical trade-offs, mechanisms, and practical limitations (Hort et al., 2022, Mahamadou et al., 22 Oct 2024, Alloula et al., 27 May 2025).

1. Taxonomy and Core Mechanisms of Bias Mitigation

Bias mitigation methods are conventionally partitioned according to the intervention point in the ML lifecycle:

Pre-processing methods operate directly on training data, reweighting or transforming samples to simulate a fairer distribution. Examples include class or subgroup reweighting, generative data augmentation (e.g. targeted counterfactual generation), or mapping features into representation spaces that obscure protected attributes (Hort et al., 2022, Mikołajczyk-Bareła et al., 2023, Pablo et al., 2023).
In-processing methods modify model objectives or training dynamics, introducing fairness constraints, adversarial debiasing losses, or dependence-minimizing regularizers. These address bias during model fitting and can enforce various statistical or equalized error-rate constraints (Ganesh et al., 17 Nov 2024, Han et al., 2021, Shrestha et al., 2021). Adversarial debiasing, mutual information penalization, and group DRO (distributionally robust optimization) paradigms exemplify this class.
Post-processing methods adjust predictions or decision thresholds after model training. This includes group-specific threshold calibration (e.g. equalized odds adjustment), randomized decision rules, or recalibration techniques ensuring multicalibration/multiaccuracy on held-out data (Mahamadou et al., 22 Oct 2024, Pablo et al., 2023).

The selection between these approaches is constrained by model access, legal context, and available data. In-processing provides the strongest parity guarantees when model weights and protected-attribute labels are available; pre-processing supports deployment-agnostic pipelines; post-processing is suitable for unmodifiable or black-box models (Hort et al., 2022, Mahamadou et al., 22 Oct 2024).

2. Fairness Formalisms and Metrics

Bias mitigation is typically oriented toward achieving parity under group fairness metrics, although predictive parity, calibration, and causal fairness are also targets.

Demographic Parity (DP): Requires $\Pr(\hat Y=1|A=0) = \Pr(\hat Y=1|A=1)$ . Operationalized as DP gap $|\Pr(\hat Y=1|A=0) - \Pr(\hat Y=1|A=1)|$ . Enforced via reweighing, threshold selection, or regularizers (Hort et al., 2022, Ganesh et al., 17 Nov 2024).
Equalized Odds (EO): Seeks to equate true positive and false positive rates across groups: $|\mathrm{TPR}_0 - \mathrm{TPR}_1|$ and $|\mathrm{FPR}_0 - \mathrm{FPR}_1|$ (Ganesh et al., 17 Nov 2024).
Equal Opportunity (EOp): A restricted form of EO, focusing solely on aligning TPRs (Ganesh et al., 17 Nov 2024, Han et al., 2021).
Predictive Parity: Equates positive predictive value (PPV) across groups.
Counterfactual Fairness: Requires invariance under hypothetical interventions on protected attributes, defined by $P(\hat Y_{A\leftarrow a}(U) = y | X=x, A=a) = P(\hat Y_{A\leftarrow a'}(U) = y | X=x, A=a)$ (Mahamadou et al., 22 Oct 2024).

Empirical studies utilize these metrics in tandem with overall accuracy or AUC to characterize the fairness–performance trade-off, and recommend routine multi-metric reporting to reveal unanticipated disparities or statistical impossibility phenomena (Hort et al., 2022, Chand et al., 23 Nov 2025, Favier et al., 21 Mar 2024).

3. Algorithmic Techniques and Recent Variations

Pre-processing

Reweighting and Resampling: Classical analysis prescribes weights $w(a, y) = \frac{P(A=a) P(Y=y)}{P(A=a, Y=y)}$ to achieve group-conditional independent sampling (Hort et al., 2022, Mahamadou et al., 22 Oct 2024). Resampling matches subgroup distributions at batch or epoch level (Pablo et al., 2023).
Targeted Data Augmentation: Controlled introduction of counterfactual and spurious examples (e.g., via mask insertion, image overlays) immunizes models to specific detectable biases and spurious features (Mikołajczyk-Bareła et al., 2023).
Proxy-Label Generation: In contexts lacking explicit protected-attribute labels, unsupervised embedding and clustering recovers proxy labels that, when integrated into downstream debiasing modules (Fair Mixup, adversarial debiasing), yield comparable fairness to that obtained using true sensitive attributes (Chaudhary et al., 2023).

In-processing

Adversarial Debiasing: Minimize predictive loss under adversary pressure to remove protected-attribute signal, typically via alternating or gradient-reversal optimization (Ganesh et al., 17 Nov 2024, Mahamadou et al., 22 Oct 2024, Hort et al., 2022).
Fairness-Constrained Risk Minimization: Augment empirical risk with fairness constraints or regularizers corresponding to chosen group metrics (DP, EO) (Ganesh et al., 17 Nov 2024, Hort et al., 2022). Examples include difference-of-probabilities penalties, mutual information penalization (Prejudice Remover), and kernel-based independence regularization (HSIC) (Ganesh et al., 17 Nov 2024).
Group DRO: Minimize worst-group loss, operationalized by maximizing group-specific empirical risk during training (Alloula et al., 27 May 2025).
Assumption-Free Interaction Modeling: Methods such as FairInt model and penalize biased pseudo-sensitive-non-sensitive feature interactions, without requiring explicit protected attributes at inference time (Chang et al., 2023).

Post-processing

Threshold Adjustment: Learning group-specific or global thresholds that minimize DP or EO gaps under an LP or search over confusion matrix entries (Pablo et al., 2023, Mahamadou et al., 22 Oct 2024).
Calibration and Multicalibration: Black-box recalibration of outputs to equalize predictive reliability across many subgroups or slices (Mahamadou et al., 22 Oct 2024).
Decoupled Inference: Training and fusing per-group classifiers or domain experts reduces bias amplification by isolating spurious group-target correlations and combining predictions at score or logit level (Wang et al., 2019).
Prompt-Based and Geometric Interventions (LLMs): In LLMs, logit steering, multi-layer projection, bias-edit parameter editing, and prompt-based guidance each shift predictions away from targeted bias axes but can induce collateral damage on untargeted dimensions, as shown on the StereoSet benchmark (Chand et al., 23 Nov 2025).

Debiasing Word Embeddings

Geometric projection (LP, HD, INLP, OSCaR) remains foundational for rapidly neutralizing bias subspaces in static embeddings (Rathore et al., 2021, Vargas et al., 2020). Empirical evidence affirms the linear subspace hypothesis: most gender bias is captured by a single direction (Vargas et al., 2020).

4. Subgroup Definition and Robustness

The impact of subgroup selection is theoretically and empirically decisive: error and bias mitigation hinge on subgroup definitions aligning with underlying spurious mechanisms. Mitigating disparities on coarse or mismatched subgroups can paradoxically worsen fairness with respect to the intended groups, even when observed disparities are large (Alloula et al., 27 May 2025). Intersectional or class-conditional subgroups generally yield optimal reweighting or DRO performance by aligning the training distribution with the unbiased target via minimal divergence (KL) (Alloula et al., 27 May 2025).

Groupings that do not correspond to the spurious or causal drivers of disparity—such as uncorrelated or random subgroups—may even degrade test-time robustness relative to no mitigation. Empirically, fine-grained subgrouping is benign or beneficial, while moderate annotation noise is tolerated, but substantial noise impairs performance (Alloula et al., 27 May 2025). The theoretical analysis demonstrates that subgroup weights optimizing the KL divergence between train and test (unbiased) joint distributions correlate tightly with post-mitigation generalization and fairness.

5. Effectiveness, Limitations, and Implementation Trade-offs

The effectiveness of bias mitigation is circumscribed by the type of data bias (label vs. selection), the mismatch between bias mechanisms and fairness targets, and the completeness of subgroup specification:

Label Bias: For label noise or corruption conditional on (A, Y), minimization of group-fairness measures (e.g., DP gap) on the observed data recovers a fair model in the underlying true population. All in-processing and post-processing techniques targeting DP or EO are thus theoretically sound in this regime (Favier et al., 21 Mar 2024).
Selection Bias: If data collection or curation induces selection bias interdependent on group and outcome, naive fairness regularization on the observed distribution will not guarantee restored fairness, and can even degrade performance under the true, unbiased population (Favier et al., 21 Mar 2024). In these cases, semi-supervised, inverse-propensity weighting, or sample-reweighting based on the selection mechanism are necessary.
Sensitivity to Hyperparameters: Benchmarking across multiple pipelines and hyperparameter assignments frequently reveals that state-of-the-art methods do not demonstrate clear superiority under optimal tuning; performance differences are often driven more by pipeline choices than algorithmic details (Ganesh et al., 17 Nov 2024). Thus, best practice calls for Pareto-frontier reporting and evaluation across diverse configurations.
Scalability to Many Bias Variables: Explicit methods (group upweighting, DRO, IRMv1) are highly sensitive to the number and definition of bias variables, with performance collapsing under numerous or complex group structures. Implicit approaches (e.g., learning from failure, spectral decoupling) retain greater robustness but must be carefully tuned (Shrestha et al., 2021).
Collateral Damage and Spillovers: In intersectional and high-dimensional settings, targeted bias mitigation along one axis may exacerbate other axes or degrade model coherence. No intervention is universally safe, necessitating systematic multi-attribute and multi-metric auditing (Chand et al., 23 Nov 2025, Mahamadou et al., 22 Oct 2024).
Practical Dimensions in Healthcare and Societal AI: Real-world deployments require participatory definition of fairness, dynamic monitoring of population shifts, and harmonization with legal, cultural, and domain-specific constraints. Value-sensitive AI frameworks institutionalize stakeholder engagement and dynamic technical–ethical alignment (Mahamadou et al., 22 Oct 2024). Intersectional fairness requires richer validation data and risks of overfitting small subgroups.

6. Current Directions, Open Challenges, and Recommendations

Despite significant empirical and algorithmic advances, bias mitigation remains brittle and context-dependent:

Rigorous evaluation protocols and stratified benchmarks (e.g., BiasedMNIST, StereoSet) have revealed hidden biases, the brittleness of best-known approaches, and the limits of transfer across domains (Shrestha et al., 2021, Chand et al., 23 Nov 2025).
Unsupervised and assumption-free debiasing—proxy label generators, feature interaction modeling—are critical for settings lacking or restricting access to sensitive attributes (Chaudhary et al., 2023, Chang et al., 2023).
Fine-tuning mitigation for intersectional and real-world subgroup structures is an open challenge, as is post-hoc auditing for deficit or harm along untargeted group axes (Alloula et al., 27 May 2025, Chand et al., 23 Nov 2025).
Actionable Recommendations:
- If subgroup annotations are available, prioritize mitigations (reweighting, DRO, rep-fusion) over high-divergence subgroups $(A,Y)$ or intersectional cells, not simply those with measured disparities.
- Integrate routine multi-metric, multi-dataset, and multi-hyperparameter evaluations; always report certification or confidence intervals across seeds and splitting pipelines (Ganesh et al., 17 Nov 2024).
- Where sensitive attributes are unavailable, apply principled proxy or unsupervised feature interaction techniques and validate via third-party fairness diagnostics (Chaudhary et al., 2023, Chang et al., 2023).
- In settings with combined selection and label bias, apply domain adaptation, propensity reweighting, or hybrid semi-supervised protocols—mere in-processing or post-processing is insufficient (Favier et al., 21 Mar 2024).
- Proactively involve domain stakeholders in fairness definition, metric choice, and subgroup specification—especially in healthcare, finance, and other high-stakes deployments (Mahamadou et al., 22 Oct 2024).

The consensus from contemporary research is that bias mitigation efficacy is fundamentally constrained by the adequacy of subgroup specification, correct diagnosis of bias type, pipeline hyperparameter tuning, and continuous, intersectional auditing. Deployment of bias mitigation needs to be embedded in a broader, participatory, and context-sensitive AI governance process to ensure model equity and reliability (Mahamadou et al., 22 Oct 2024, Alloula et al., 27 May 2025, Chand et al., 23 Nov 2025).