Pre-processing Bias Mitigation

Reweighting adjusts the importance (weight) of individual training samples to balance the distribution of outcomes across protected groups. It is a statistical correction applied before training.

• Mechanism: Samples from underrepresented groups that receive favorable outcomes are assigned higher weights, while overrepresented groups with unfavorable outcomes may be down-weighted.
• Goal: To create a training set where the target label is statistically independent of the protected attribute, satisfying fairness criteria like demographic parity at the data level.
• Example: In a hiring dataset where "female" and "hired" are negatively correlated, reweighting increases the influence of resumes from hired females and non-hired males during model training.

Disparate Impact Removal is a pre-processing algorithm that transforms the features in a dataset to remove any information that could lead to discriminatory outcomes, as measured by the disparate impact ratio.

• Core Technique: It learns a linear transformation of the feature space to maximize utility (predictive power for the true label) while minimizing the classifier's ability to predict the protected attribute from the transformed data.
• Mathematical Goal: Achieve $P(\hat{Y} | A=0) / P(\hat{Y} | A=1) \approx 1$, where $\hat{Y}$ is the prediction and $A$ is the protected attribute, by manipulating the input features $X$.
• Outcome: The processed data $X'$ can be used with any standard classifier, as the bias mitigation is baked into the features.

Learning Fair Representations (LFR) is an optimization-based technique that maps the original data into a new, latent representation designed to obfuscate protected group membership while preserving utility for the main task.

• Process: An encoder network learns to produce a representation $Z$ from input $X$. The training objective has three competing terms:
  • Reconstruction Loss: $Z$ should allow decoding back to something similar to $X$.
  • Adversarial Loss: A critic cannot accurately predict the protected attribute $A$ from $Z$.
  • Task Loss: $Z$ should be predictive of the true label $Y$.
• Advantage: Produces a debiased feature set that can be used for downstream modeling, separating the fairness intervention from the final model choice.

These are foundational, often manual, techniques for altering training data to reduce bias.

• Suppression: The direct removal of protected attributes (e.g., race, gender) from the dataset. This is often legally required but is insufficient on its own due to proxy variables (e.g., zip code, name frequency) that can leak protected information.
• Label Flipping / Massaging: Selectively changing the value of the target label $Y$ for specific instances to improve fairness metrics.
  • Method: Identify instances near the decision boundary where flipping the label (e.g., from "deny" to "approve") would most improve statistical parity.
  • Limitation: This alters ground truth, which may not be ethically or legally permissible in high-stakes domains and can reduce dataset integrity.

Pre-processing is one of three intervention points in the ML pipeline, each with distinct trade-offs.

• Pre-processing (Data-Level):

  • Pros: Agnostic to model choice; addresses bias at the source.
  • Cons: Alters the fundamental training data; may reduce utility.

• In-processing (Algorithm-Level): Modifies the training objective (e.g., adding fairness constraints).

  • Pros: Can directly optimize fairness-accuracy trade-off.
  • Cons: Tied to specific model families; requires custom implementations.

• Post-processing (Output-Level): Adjusts predictions after training (e.g., applying different decision thresholds per group).

  • Pros: Simple to implement; no retraining needed.
  • Cons: Requires access to protected attributes at inference, which may be prohibited; can be seen as "fairness through blindness".

Reweighting adjusts the importance (weight) of individual training examples to balance the distribution of outcomes across protected groups. It is a foundational pre-processing technique.

• Mechanism: Calculates weights for each data point so that, in the weighted dataset, the probability of a positive label is independent of the protected attribute.
• Use Case: Corrects for historical bias where past discriminatory decisions have skewed the dataset. For example, if 'loan approval' in historical data is biased against a group, reweighting gives more importance to approved applicants from that underrepresented group.
• Effect: The model learns from a statistically fairer version of the data without altering the original feature values.

The Disparate Impact Remover is an algorithm that edits feature values to reduce discrimination while preserving rank-ordering within groups. It is implemented in toolkits like IBM's AIF360.

• Mechanism: Operates on non-protected, numeric features. It applies a massaging technique, transforming the distribution of features for the disadvantaged group to more closely match the distribution of the advantaged group.
• Objective: Achieves a target level of demographic parity (statistical parity) in the repaired dataset.
• Consideration: This is a transformative method. It changes the underlying data, which can be desirable for fairness but may reduce utility if applied too aggressively.

Learning Fair Representations (LFR) is a pre-processing technique that learns a new, encoded representation (Z) of the data that obfuscates information about protected attributes while retaining utility for the prediction task.

• Mechanism: Uses an optimization framework with three competing objectives: 1) Reconstruction loss (Z should allow reconstruction of original non-protected features), 2) Prediction loss (Z should be useful for predicting the target label Y), and 3) Adversarial loss (Z should prevent prediction of the protected attribute A).
• Output: A transformed, fairness-aware dataset (in the Z-space) used for subsequent model training.
• Advantage: Provides a strong separation between the learned representations and sensitive attributes, enabling fairness through obscurity.

Optimized Pre-processing formulates bias mitigation as a convex optimization problem to find the closest possible fair dataset to the original data, where closeness is measured by probability distributions.

• Mechanism: Given original distributions P(X, A, Y), it finds new distributions Q(X, A, Y) that satisfy selected group fairness constraints (like demographic parity or equalized odds) while minimizing the Wasserstein distance or KL-divergence between P and Q.
• Result: Produces a transformed dataset with modified labels and/or features. In practice, this often results in label flipping for select instances to meet the fairness goal.
• Guarantee: Provides a theoretically grounded, optimal transformation for the specified fairness metric and distance measure.

Algorithmic fairness is the study and application of principles and techniques to ensure that automated decision-making systems do not create or perpetuate unjust or discriminatory outcomes against individuals or groups based on sensitive attributes. It is the overarching goal that pre-processing techniques aim to support.

• Core Concern: Preventing harm from automated decisions in areas like hiring, lending, and criminal justice.
• Technical Challenge: Formalizing often competing definitions of "fairness" (e.g., demographic parity vs. equal opportunity) into measurable objectives.
• Trade-offs: Often involves balancing fairness metrics with overall model accuracy and utility.

Bias in data refers to systematic skews or inaccuracies in a dataset that can lead a model trained on that data to produce unfair or inaccurate outputs. Pre-processing techniques directly target these data-level issues.

• Historical Bias: Arises when past societal inequities are captured in training data (e.g., historical hiring data favoring one demographic).
• Representation Bias: Occurs when the dataset does not adequately represent the diversity of the target population.
• Measurement Bias: Introduced by flawed data collection instruments or procedures.
• Aggregation Bias: Happens when data from diverse groups is inappropriately combined, masking important subgroup differences.

A protected attribute is a personal characteristic, such as race, gender, age, religion, or disability status, that is legally or ethically prohibited from being used as a basis for discriminatory treatment in algorithmic decision-making.

• Role in Pre-processing: These attributes are the central axis for identifying and measuring bias. Techniques often aim to decorrelate other features from these attributes or reweight data based on them.
• Explicit vs. Proxy Exclusion: While protected attributes are often removed from training data, proxy variables (e.g., zip code for race) can still enable discrimination, which pre-processing must also address.
• Jurisdictional Variation: The specific list of protected attributes can vary by country (e.g., the EU's AI Act, US Civil Rights Act).

A fairness metric is a quantitative measure used to assess whether an AI model's performance or predictions are equitable across different demographic subgroups defined by protected attributes. These metrics define the target for pre-processing interventions.

• Demographic Parity: Requires the overall rate of positive predictions (e.g., loan approvals) to be equal across groups.
• Equal Opportunity: Requires the true positive rate (recall) to be equal across groups.
• Equalized Odds: A stricter condition requiring both true positive rates and false positive rates to be equal across groups.
• Selection: The choice of metric involves ethical and legal considerations and dictates the appropriate mitigation strategy.

In-processing bias mitigation involves techniques applied during model training to directly optimize for both accuracy and fairness, contrasting with pre-processing's focus on data.

• Fairness Constraints: Mathematical conditions (e.g., demographic parity) are added directly to the model's optimization objective.
• Adversarial Debiasing: A primary model is trained to make accurate predictions while an adversarial network tries to predict the protected attribute from the primary model's internal representations, forcing them to be uninformative about the attribute.
• Trade-off: Offers more direct control over the learning objective but requires modifying the core training loop, which can be more complex than pre- or post-processing.

Post-processing bias mitigation involves techniques applied to a model's predictions after training to achieve a desired fairness metric, without retraining the model.

• Threshold Adjustment: Different decision thresholds are applied to the model's score outputs for different demographic groups to equalize error rates (e.g., to achieve Equalized Odds).
• Advantage: Simple to implement and deploy, as it treats the model as a fixed black box.
• Limitation: Does not address root causes of bias in the model or data and can sometimes reduce overall utility. It is often used when model retraining is prohibitively expensive or impossible.