Decoupling Fairness from Model Architecture through Data Repair

This direction is split into other directions: model-agnostic fairness pre-processing. Its last version stays readable and citable here.

Version 1 of 172 papers94 conceptsfrozen 2026-10-04

In the effort to build equitable AI, researchers typically intervene at one of three stages: pre-processing (cleaning the training data), in-processing (modifying the model's loss function), or post-processing (adjusting the model's final predictions). While in-processing is popular, it tethers fairness to a specific model architecture. If you switch from a linear regression to a deep neural network, you must redesign your fairness constraints from scratch.

A growing body of research suggests a more portable alternative: treating fairness as a property of the data distribution itself. By "repairing" the underlying data before it ever reaches a learner, we can ensure that any downstream model—regardless of its complexity—inherits these fairness guarantees.

The Intuition of Data Repair

The core philosophy of this direction, as established in the foundational work Certifying and Removing Disparate Impact by Michael Feldman, Sorelle A. Friedler, and colleagues, is that bias is often a form of "information leakage" in the dataset. If a protected attribute (like gender or race) is statistically or causally linked to the target label, a high-performance model will inevitably exploit that link.

Pre-processing works by "scrubbing" this link. If the training data is transformed such that the protected attribute and the target are independent, a model trained on that data cannot discriminate because the discriminatory signal has been removed from the source. This creates a "model-agnostic" workflow: once the data is repaired, practitioners can use off-the-shelf, high-performance black-box algorithms without further modification.

Why Pre-Processing Outperforms Post-Processing

It might seem easier to simply tweak a model's outputs to be fair (post-processing). However, Sajad Khodadadian, AmirEmad Ghassami, and Negar Kiyavash mathematically demonstrate in Impact of Data Processing on Fairness in Supervised Learning that pre-processing is fundamentally more effective.

They show that a pre-processor can replicate the results of any post-processor while achieving a superior trade-off between fairness and accuracy. This is because post-processing is a "blind" adjustment that doesn't account for the features that led to the prediction, whereas pre-processing allows the model to learn the most accurate patterns possible from a sanitized foundation.

Mechanisms for Sanitizing Data

Researchers have developed several distinct mechanisms to achieve this "sanitization" across different data types:

1. Geometric and Distributional Alignment

For tabular data, the goal is often to transform feature distributions so they look identical across different demographic groups. Flavio Calmon and Kush R. Varshney introduced a convex optimization framework in Optimized Pre-Processing for Discrimination Prevention that learns a randomized transformation. This transformation balances three goals: reducing group discrimination, limiting individual distortion, and preserving predictive utility. More recently, Shixiang Zhu and Shuyi Chen proposed Orthogonal to Bias (OB), which uses closed-form geometric projections to make non-sensitive data orthogonal to sensitive attributes, offering a significant speedup over iterative neural methods.

2. Strategic Label Flipping

Sometimes, the bias isn't in the features, but in the labels themselves (e.g., historical hiring bias). Frameworks like iFlipper by Hantian Zhang and colleagues strategically flip a minimal number of training labels to satisfy individual fairness. Similarly, FairRR uses a randomized response framework to modify labels, allowing practitioners to precisely tune the level of disparity they are willing to tolerate.

3. Structural Repair in Graphs

In network data, bias often propagates through the graph's structure (homophily). Ievgen Redko and Charlotte Laclau address this in All of the Fairness for Edge Prediction with Optimal Transport by using Optimal Transport to "repair" the graph's adjacency matrix. By aligning the distributions of connections across different groups before any embedding or prediction occurs, they ensure that the resulting graph is "fair-by-design."

4. Generative Balancing

Rather than deleting or modifying existing data, which can lead to information loss, some researchers propose adding synthetic data. Prasanna Sattigeri and Kush R. Varshney introduced Fairness GAN to generate fair synthetic versions of biased datasets. In the medical domain, Brandon Theodorou and colleagues developed FairPlay, which uses Large Language Models to generate realistic patient data, filling demographic gaps to improve both equity and overall model performance.

Causal and Counterfactual Fairness

A sophisticated branch of this research moves beyond simple statistical correlation to causal inference. Drago Plecko and Nicolai Meinshausen developed fairadapt, which uses counterfactual reasoning to "warp" data into a version where protected attributes have no causal influence on the outcome.

This approach is particularly powerful because it addresses "proxy" variables—features that aren't protected themselves but are causally downstream of a protected attribute (like neighborhood as a proxy for race). By transforming the data to be "causally fair," researchers like Haoyu Chen and Rui Song in FLAP ensure that any model trained on the data will ignore these discriminatory pathways.

The Path Toward Fair-by-Design Data

The ultimate goal of this research direction is the creation of provably fair representations. As Daniel McNamara and Cheng Soon Ong argue in Provably Fair Representations, we can mathematically bound how much a downstream model can discriminate by controlling the information content of the data representation.

This allows for a clean separation of concerns: data producers can focus on generating high-quality, de-biased datasets, while data consumers (machine learning engineers) can focus on maximizing predictive performance, confident that the ethical constraints are already baked into the foundation of their work.

Papers in this version

Showing 60 of 72 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_decoupling_fairness_from_model_architect_f1f7d407,
  title = {Decoupling Fairness from Model Architecture through Data Repair},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decoupling-fairness-from-model-architecture-through-data-repair/rdr_f1f7d407af884ac0ba24192fa7adc241},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_decoupling_fairness_from_model_architect_f1f7d407_v1,
  title = {Decoupling Fairness from Model Architecture through Data Repair},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/decoupling-fairness-from-model-architecture-through-data-repair/rdr_f1f7d407af884ac0ba24192fa7adc241/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}