Machine learning models are inherently conservative: they assume the future will look like the past. When trained using standard optimization objectives, algorithms excel at finding and exploiting statistical patterns within historical data. However, if that data reflects systemic inequalities—such as historical underperformance of certain student demographics due to external factors, biased hiring decisions, or uneven media representation—the model treats these disparities as valid predictive signals.
As Solon Barocas et al. observe in Big Data, Data Science, and Civil Rights, standard training algorithms mistake indicators of past discrimination for predictive utility. This mathematical reliance on past patterns creates an inherent resistance to social innovation, effectively freezing existing social structures (Claude Draude et al. in Social Machines).
To break this feedback loop, researchers have developed techniques across three distinct stages of the machine learning pipeline: data pre-processing, representation learning (in-processing), and post-hoc structural adjustments.
Pre-Processing: Cleaning the Mirror
If training labels are corrupted by historical bias, we can model this corruption mathematically rather than treating the data as absolute ground truth. The core intuition is to view observed historical decisions as a "noisy proxy" for an underlying, fair decision.
In Identifying and Correcting Label Bias in Machine Learning, Heinrich Jiang et al. demonstrate that if we assume bias arises from a process that minimizes information divergence from the truth while satisfying fairness constraints, we can derive an iterative re-weighting procedure to recover an unbiased classifier without ever seeing the true labels. Similarly, YooJung Choi leverages probabilistic circuits in Group Fairness by Probabilistic Modeling with Latent Fair Decisions to explicitly model the bias mechanism, enforcing demographic parity while preserving the analytical utility of the data distribution.
Biased Historical Data ───► [Causality/Noise Model] ───► Repaired/Unbiased Data
(Labels reflect past (Isolates and decouples (Ready for standard
discrimination) sensitive attributes) model training)
Instead of re-weighting, we can also strategically alter the data. Joshua John Ward et al. introduce FairRR: Pre-Processing for Group Fairness through Randomized Response, which uses a randomized response framework to strategically "flip" labels in the training set. This breaks the statistical dependency between sensitive attributes and target labels, allowing practitioners to precisely tune the trade-off between fairness and utility.
Alternatively, generative models can synthesize entirely new datasets free from historical bias. Frameworks like those proposed by Depeng Xu et al. (FairGAN: Fairness-aware Generative Adversarial Networks) and Amirarsalan Rajabi et al. (TabFairGAN: Fair Tabular Data Generation with Generative Adversarial Networks) use a dual-discriminator setup to ensure that sensitive attributes are statistically decoupled from outcomes in the generated data.
In-Processing: Geometric and Adversarial Separation
When the training data cannot be modified, the model's internal representations must be constrained so it cannot "see" or exploit historical biases. This is achieved by separating target-relevant information from sensitive attributes within the model's latent space.
Mhd Hasan Sarhan et al. achieve this geometrically in Fairness by Learning Orthogonal Disentangled Representations by forcing the network to project data into two perpendicular (orthogonal) vector spaces:
$$\mathbf{z} = \mathbf{z}_{\text{task}} \perp \mathbf{z}_{\text{sensitive}}$$
By maximizing the entropy of the sensitive space, the network actively strips out discriminatory correlations while retaining the information necessary for the task. In a similar vein, Enze Shi et al. introduce Sequential Fair Projection in Understanding Fairness and Prediction Error through Subspace Decomposition and Influence Analysis, which decomposes features into target-relevant, sensitive, and shared components, allowing developers to filter out the shared components at the root representation level.
Rather than enforcing geometric constraints, we can introduce an active opponent during training. Brian Hu Zhang et al. (Mitigating Unwanted Biases with Adversarial Learning) set up an adversarial framework where a predictor network learns to perform the task, while an adversary network simultaneously tries to guess protected attributes from the predictor's outputs. By penalizing the predictor whenever the adversary succeeds, the model is forced to minimize the propagation of historical bias.
┌───► Predictor ───► Task Output
│
Input Data ───► Latent Space
│
└───► Adversary ───► Protected Attribute (Penalizes Predictor)
Post-Processing: Structural and Causal Repairs
Retraining large models from scratch is often computationally prohibitive or impossible due to data access limits. Post-processing techniques accept a biased model as given and adjust its internal structure or final decisions to satisfy fairness criteria.
Using causal inference, Yixin Wang et al. demonstrate in Adjusting Machine Learning Decisions for Equal Opportunity and Counterfactual Fairness that predictions can be adjusted post-hoc to satisfy counterfactual fairness criteria. By leveraging a causal graph of the decision environment, their method maintains high fidelity to the original predictions while stripping away historical biases embedded in downstream variables.
Alternatively, the internal logic of the model can be surgically repaired. Jie Zhang et al. challenge the assumption that fairness must come at the cost of accuracy in Search-based Automatic Repair for Fairness and Accuracy in Decision-making Software. They use search-based software engineering to mutate specific structural components—such as decision tree branches—that have captured discriminatory correlations from historical data. This targeted mutation allows them to improve both fairness and accuracy simultaneously, proving that models can be structurally repaired to prevent the automated replication of historical bias.
Key Foundational Work
Discrimination Aware Decision Tree Learning (2010) — contributed 1 concept elements to this direction.
Discrimination-Aware Classifiers for Student Performance Prediction (2015) — contributed 1 concept elements to this direction.