Bias-Aware Training Techniques

• Bias-aware training is a set of approaches that modifies the learning process to account for systematic distortions in data, labels, and annotations.
• It employs specialized methods such as adversarial loss adjustments, contrastive representation shaping, and sample reweighting to counter various bias types.
• These techniques improve model fairness, robustness, and sample efficiency by aligning training objectives with fairness and deployment criteria.

Bias aware training is a family of learning procedures in which model fitting is modified to account for systematic distortions in data, labels, annotations, or evaluation targets rather than treating them as ordinary stochastic noise. In the literature covered here, those distortions include unequal true-positive rates across protected groups, label bias and selection bias in historical decision datasets, dataset-specific shifts in Mean Opinion Scores (MOS), presentation bias in recommendation logs, spurious class–attribute correlations in visual data, prior shift between training and deployment prevalence, and systematic error in low-fidelity scientific labels (Han et al., 2022, Verma et al., 2021, Mittag et al., 2021, Yi et al., 2021, Folgoc et al., 2020, Bukharin et al., 2023). The resulting methods span objective design, adversarial learning, contrastive representation shaping, sample reweighting, pseudo-label selection, data editing, and multi-stage fine-tuning.

1. Sources of bias and formal problem settings

A central distinction in this literature is between biases defined at the level of model error rates and biases defined at the level of data generation. In fairness-oriented classification, the concern is often group-conditional performance disparity: a model may be accurate on average while exhibiting substantially different true-positive rates across protected groups. In historical decision data, bias may originate from biased labels or from biased selection into the observed dataset. Verma et al. explicitly distinguish label bias and selection bias, and evaluate both individual discrimination and Statistical Parity Difference (SPD); under their formulation, similar individuals are defined via a distance $d_x(\cdot,\cdot)$ that ignores the sensitive attribute, and individual discrimination is the fraction of similar pairs assigned different predictions (Verma et al., 2021).

A second class of formulations treats bias as distribution shift. Under sampling bias or prior shift, the class prior in the training set, $p_{\mathrm{train}}(y)$, differs from the deployment prior, $p_{\mathrm{test}}(y)$, so a standard cross-entropy model converges to the Bayes-optimal classifier for the training distribution rather than for the test distribution (Folgoc et al., 2020). Du and Wu instead formulate sample selection bias through a selection indicator $s \in \{0,1\}$ and a Missing-At-Random assumption $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$, which induces instancewise density-ratio corrections and a worst-case fairness problem under uncertainty in those ratios (Du et al., 2021).

A third class concerns structured nuisance factors that systematically alter labels or clicks. In blind quality prediction from multiple subjective datasets, the same file may receive different MOS values across experiments because of rating noise, order-effect, corpus-effect, and long-term dependencies; the paper models these as dataset-specific offset and gain distortions (Mittag et al., 2021). In personalized news recommendation, click logs are biased by presentation variables such as position and size, so historical clicks are not treated as a direct proxy for user preference (Yi et al., 2021). In Web-sourced image classification, bias may appear as a class-specific attribute that is over-represented for one class; BiasEdit formalizes this through statistical dependence and mutual information between labels and attributes extracted by a pretrained vision–LLM (Seo et al., 27 May 2026).

These formulations imply different intervention points. Some methods alter the loss so that the training objective matches a fairness criterion or deployment distribution. Others alter the representation so protected attributes or nuisance factors are harder to recover. Still others alter the data itself by reweighting, subsampling, editing, or pseudo-label selection. This suggests that “bias aware” is not a single algorithmic template but a design principle tied to the causal or statistical form of the bias under study.

2. Objectives aligned with fairness criteria

A recurrent theme is the mismatch between fairness motivations and standard training objectives. For equal opportunity, the criterion is class-conditional equality of true-positive rates. Given inputs $X$, labels $Y \in \{1,\dots,|C|\}$, protected attribute $G \in \mathcal G$, and prediction $\hat Y$, Han et al. define

$$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$

They then set $p_{\mathrm{train}}(y)$0, aggregate over groups as $p_{\mathrm{train}}(y)$1, and define

$p_{\mathrm{train}}(y)$2

A perfectly equal-opportunity model satisfies $p_{\mathrm{train}}(y)$3, and the paper reports “fairness” as $p_{\mathrm{train}}(y)$4 (Han et al., 2022).

A more direct objective-level treatment is given by Savani et al., who define subgroup- and class-specific cross-entropies $p_{\mathrm{train}}(y)$5 and $p_{\mathrm{train}}(y)$6 and augment the ordinary batch loss with penalties that encourage those terms to align. Their class-wise equal-opportunity objective is

$p_{\mathrm{train}}(y)$7

while $p_{\mathrm{train}}(y)$8 adds an inter-class alignment term based on $p_{\mathrm{train}}(y)$9. The paper’s theoretical justification is that, in the binary case, $p_{\mathrm{test}}(y)$0 approximates $p_{\mathrm{test}}(y)$1, so minimizing it drives $p_{\mathrm{test}}(y)$2 (Shen et al., 2022).

Balanced training offers a different proxy. In “Balancing out Bias,” the equal-opportunity objective is implemented by reweighting examples according to the joint distribution of class and protected attribute. In the multi-class case,

$p_{\mathrm{test}}(y)$3

and, at the instance level, each example receives

$p_{\mathrm{test}}(y)$4

This makes the learning procedure explicitly aware of which $p_{\mathrm{test}}(y)$5 cells are under- or over-represented, and the paper also studies down-sampling to a balanced dataset as an alternative realization of the same objective (Han et al., 2021).

Across these approaches, the common move is to replace generic empirical risk minimization by a criterion whose terms can be interpreted in the same conditional regime as the fairness target. This is especially important for equal opportunity, where unconditional invariance can misalign with the intended notion of fairness.

3. Adversarial, contrastive, and architecture-level mechanisms

Adversarial training remains one of the standard templates for debiasing, but its target depends on what the discriminator is asked to recover. In the standard encoder–classifier–discriminator setup, the task loss is

$p_{\mathrm{test}}(y)$6

the discriminator loss is

$p_{\mathrm{test}}(y)$7

and the minimax problem is

$p_{\mathrm{test}}(y)$8

Han et al. argue that this standard form enforces $p_{\mathrm{test}}(y)$9, which corresponds to demographic parity, whereas equal opportunity requires $s \in \{0,1\}$0. Their augmented discriminator therefore conditions on the true class via an augmentation layer

$s \in \{0,1\}$1

with shared and class-specific projectors. The discriminator then approximates $s \in \{0,1\}$2, so adversarial unlearning explicitly targets class-conditional bias (Han et al., 2022).

Contrastive learning provides a non-adversarial representation-level alternative. In “Contrastive Learning for Fair Representations,” representations of examples sharing the same task label are pulled together, while representations of examples sharing the same protected attribute are pushed apart. The combined objective is

$s \in \{0,1\}$3

where $s \in \{0,1\}$4 is a supervised contrastive term over the main label and $s \in \{0,1\}$5 is the analogous term over the protected attribute. The stated effect is to keep label information in focused regions while ensuring the protected attribute has a diverse spread. The paper reports compute overhead of $s \in \{0,1\}$6–$s \in \{0,1\}$7 relative to vanilla cross-entropy, compared with $s \in \{0,1\}$8–$s \in \{0,1\}$9 slower adversarial baselines (Shen et al., 2021).

Other mechanisms alter the architecture so nuisance factors are explicitly modeled rather than merely suppressed. The Distraction module defines a two-player minimax game in which a submodule is trained through a fairness loss that predicts the sensitive attribute from the model’s own score, while the classifier is optimized for the main task. The final objective is

$P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$0

and the method is instantiated on tabular, graph, and vision models (Yazdani-Jahromi et al., 2022). DebiasRec, in turn, decomposes click score into a preference score $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$1 and a bias score $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$2, and uses only the preference component for ranking at inference time, while user modeling itself is calibrated by bias embeddings derived from position and size (Yi et al., 2021).

A related but more localized intervention appears in NER. To mitigate Name Regularity Bias, Ghaddar et al. add learnable adversarial noise to entity mentions through a noise-embedding table indexed by corrupted entity types, train an auxiliary head on noisy labels, and combine this with masking and parameter freezing. The perturbation is injected at the input-embedding level, so the method is model-agnostic across BERT-based and LSTM-CRF NER systems (Ghaddar et al., 2021).

4. Reweighting, resampling, and data-centric correction

One major branch of bias-aware training estimates which samples, labels, or subsets are unreliable and adjusts their influence accordingly. In ASTEROID, the setting is a large inaccurate dataset $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$3 and a small accurate dataset $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$4. A temporary model trained on $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$5 produces a per-example bias surrogate

$P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$6

which is converted into a weight

$P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$7

The inaccurate dataset is then used through a weighted loss $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$8, after which the model is fine-tuned on $P(s=1 \mid x,a,y)=P(s=1 \mid x,a)$9. The paper also provides a score-matching variant for unlabeled inaccurate data (Bukharin et al., 2023).

Under prior shift, a simpler weighting rule suffices. Bayesian Sampling Bias Correction derives a bias-corrected loss from Bayesian risk minimization by setting

$X$0

and minimizing the weighted cross-entropy

$X$1

The paper emphasizes that minimizing this objective is equivalent to minimizing the KL divergence between the true test-conditional distribution and the model prediction under the test prior, and the same replacement can be made inside the ELBO of a Bayesian neural network (Folgoc et al., 2020).

When selection bias is instance-dependent and density-ratio estimates are uncertain, Du and Wu propose a minimax robust formulation. They define instance weights via $X$2 and optimize a worst-case reweighted risk over an $X$3-ball around the estimated selection probabilities, with fairness enforced through a weighted boundary-fairness surrogate. Two algorithms are given: RFLearn¹, when unlabeled test data are available for direct density-ratio estimation, and RFLearn², when they are not and cluster-level ratios are used instead (Du et al., 2021).

Other methods intervene directly on the training set. Verma et al. rank training points by their average influence on “victim” predictions extracted from discriminatory similar pairs, then remove the highest-ranked points until the first local minimum of individual discrimination is reached (Verma et al., 2021). TAB trains a helper ERM model, records per-sample loss histories $X$4, clusters them per class with $X$5-means($X$6), treats the smaller cluster as enriched in bias-conflicting samples, and upsamples it to construct a group-balanced training multiset without group labels or model selection (Zarlenga et al., 2024). DCAST extends self-training by selecting pseudo-labeled samples per class with a diversity constraint based on distances in embedding space, thereby countering confirmation bias while addressing class-aware selection bias (Tepeli et al., 2024). BiasEdit goes further upstream: it automatically detects unknown bias attributes via statistical dependence and mutual information, edits those attributes with text-guided image editing to generate bias-conflict samples, and then retrains a standard classifier by ordinary empirical risk minimization on the augmented dataset (Seo et al., 27 May 2026).

These methods differ in whether they require protected attributes, an auxiliary accurate dataset, unlabeled data from the target population, or only the original biased dataset. Their shared principle is that the effective training distribution should be altered to better represent either the intended deployment population or the bias-conflict slices that ordinary ERM underutilizes.

5. Domain-specific instantiations

Bias-aware training is not confined to protected-attribute fairness. In image and speech quality prediction, dataset-specific bias is modeled as an affine map applied to the network output for each dataset:

$X$7

leading to a bias-aware loss

$X$8

The network weights are optimized by back-propagation, while the per-dataset bias parameters are updated after each epoch by closed-form least squares once the model exceeds a Pearson correlation threshold $X$9 on the training set (Mittag et al., 2021).

In recommendation, DebiasRec decomposes the problem into a Bias Representation Module, a Bias-Aware User Modeling module, and a Bias-Aware Click Prediction module. Position and size are embedded, their interaction is modeled by element-wise product, and the resulting bias embedding both modulates historical click attention and contributes an additive bias score at training time. At inference, ranking is performed with the content-preference score only, so the deployed ranking signal is explicitly debiased with respect to the logged presentation bias (Yi et al., 2021).

In low-data adaptation from a potentially biased pretrained model, “Addressing Bias Through Ensemble Learning and Regularized Fine-Tuning” uses a counter-bias regularizer

$Y \in \{1,\dots,|C|\}$0

inside

$Y \in \{1,\dots,|C|\}$1

The small dataset is split into biased subsets emphasizing under-represented classes, local models are fine-tuned from $Y \in \{1,\dots,|C|\}$2 on those subsets, predictions are aggregated by an ensemble, and the ensemble may be distilled into a single student with a KL-based distillation loss (Radwan et al., 2024).

Scientific machine learning shows a distinct use of the same principle. ASTEROID treats cheap inaccurate labels as structurally informative but systematically biased, so bias is not removed by filtering them out entirely; instead, their contribution is smoothly down-weighted and then corrected by a small accurate set. This suggests that bias-aware training can be understood not only as fairness mitigation but also as fidelity-aware optimization under heterogeneous supervision (Bukharin et al., 2023).

6. Empirical behavior, evaluation practice, and open problems

The empirical literature evaluates bias-aware training with highly heterogeneous metrics. Fairness-oriented NLP and vision papers commonly report accuracy together with equal-opportunity gap $Y \in \{1,\dots,|C|\}$3, true-positive-rate gaps, worst-group accuracy, or leakage of protected attributes from hidden representations or logits (Han et al., 2022, Shen et al., 2021, Zarlenga et al., 2024). Historical-decision and selection-bias settings often add individual discrimination and SPD (Verma et al., 2021). Quality prediction papers use Pearson correlation coefficient, while force-field papers use MAE and MD stability (Mittag et al., 2021, Bukharin et al., 2023). This measurement diversity reflects different operational definitions of “bias,” and it limits direct cross-paper comparison.

Within equal-opportunity optimization, conditioning the debiasing mechanism on the target class yields consistent gains over unconditional adversarial removal. On Moji under a $Y \in \{1,\dots,|C|\}$4 accuracy slack, A-Adv improves fairness from $Y \in \{1,\dots,|C|\}$5 to $Y \in \{1,\dots,|C|\}$6 and reduces GAP from $Y \in \{1,\dots,|C|\}$7 to $Y \in \{1,\dots,|C|\}$8 relative to standard adversarial debiasing; on Bios, A-Adv improves fairness from $Y \in \{1,\dots,|C|\}$9 to $G \in \mathcal G$0 and reduces GAP from $G \in \mathcal G$1 to $G \in \mathcal G$2 (Han et al., 2022). Direct EO objectives also perform strongly: on Moji, $G \in \mathcal G$3 attains macro-$G \in \mathcal G$4 with $G \in \mathcal G$5, the best GAP in Table 1, and $G \in \mathcal G$6 achieves $G \in \mathcal G$7 (Shen et al., 2022). Contrastive fair representation learning reports on Moji an improvement from CE $G \in \mathcal G$8 to $G \in \mathcal G$9 in $\hat Y$0, together with lower hidden and logit leakage (Shen et al., 2021).

Data-centric approaches also show large empirical effects. Verma et al. report, for $\hat Y$1, average individual discrimination dropping from $\hat Y$2 to $\hat Y$3, test accuracy rising from $\hat Y$4 to $\hat Y$5, and SPD falling from $\hat Y$6 to $\hat Y$7 when training on the debiased subset rather than the full historical data (Verma et al., 2021). TAB improves worst-group accuracy selected by validation accuracy on all listed tasks, for example from $\hat Y$8 to $\hat Y$9 on cMNIST and from $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$0 to $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$1 on Waterbirds, while preserving or improving mean accuracy (Zarlenga et al., 2024). DCAST-10 significantly outperforms supervised-biased baselines on all $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$2 datasets with neural networks, with average $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$3 accuracy, and DCAST-100 adds a $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$4 improvement over supervised-biased training on MNIST (Tepeli et al., 2024). BiasEdit reports Avg accuracy of approximately $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$5 on Colored MNIST, $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$6 on BFFHQ, $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$7 on Dogs & Cats, and $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$8 on Waterbirds under a fully biased training setting with $$TPR_{g,y} = Pr[\hat Y = y \mid Y = y, G = g], \qquad TPR_y = Pr[\hat Y = y \mid Y = y].$$9 bias-conflict samples in the original data (Seo et al., 27 May 2026).

Outside fairness, the same pattern holds. Bias-aware MOS training raises synthetic-set Pearson CC from approximately $p_{\mathrm{train}}(y)$00 to approximately $p_{\mathrm{train}}(y)$01 when training on biased data, close to the approximately $p_{\mathrm{train}}(y)$02 obtained on an unbiased training set, and improves four of five image-quality datasets and five of six speech-quality datasets in leave-one-out evaluation (Mittag et al., 2021). ASTEROID, at a cost ratio of $p_{\mathrm{train}}(y)$03 cheap:gold, cuts test MAE by approximately $p_{\mathrm{train}}(y)$04 for GemNet and approximately $p_{\mathrm{train}}(y)$05 for EGNN relative to standard training with the same total budget; with empirical force-field data, bias-aware training reduces GemNet error by $p_{\mathrm{train}}(y)$06 and EGNN by $p_{\mathrm{train}}(y)$07 (Bukharin et al., 2023). These results indicate that bias-aware training can improve both robustness and sample efficiency when bias arises from low-fidelity supervision rather than from protected attributes.

Several limitations recur. Some methods are restricted to binary protected attributes, notably the EO objectives of Savani et al. (Shen et al., 2022). Some retain fixed encoders, leaving open whether joint fine-tuning would reduce or reintroduce bias (Shen et al., 2022). Methods based on protected attributes or logged bias variables require those signals to be available and correctly specified, which excludes unknown or unlogged biases (Yi et al., 2021, Han et al., 2021). DCAST requires unlabeled data that contain the missing support, while TAB trades hyperparameter-free deployment for two full training runs and $p_{\mathrm{train}}(y)$08 loss-history storage (Tepeli et al., 2024, Zarlenga et al., 2024). More broadly, debiasing with respect to one attribute does not guarantee the absence of bias on other attributes or intersections (Shen et al., 2022).

Taken together, the literature indicates that bias-aware training is most effective when the intervention matches the structure of the bias: class-conditional objectives for equal opportunity, dataset-specific calibration for annotation shifts, density-ratio weighting for prior or selection shift, representation conditioning for nuisance-feature suppression, and data-centric augmentation or editing when the biased support itself must be changed. This suggests that the main technical question is not whether to perform bias-aware training, but which bias model is sufficiently faithful to justify the resulting optimization.