Mitigating Indirect Discrimination via Redundant Encoding Analysis

This direction is merged into another direction: proxy variable bias (proxy-based bias). Its last version stays readable and citable here.

Version 1 of 131 papers34 conceptsfrozen 2026-10-04

The most intuitive approach to algorithmic fairness is "Fairness through Unawareness" (FtU)—simply deleting sensitive attributes like race, gender, or age from a dataset. However, a foundational body of research has demonstrated that this approach is almost always insufficient. Because real-world data is deeply interconnected, sensitive information is rarely contained in a single column; instead, it is "smeared" across seemingly neutral features.

The Illusion of Unawareness

The core problem with FtU is a phenomenon known as redundant encoding. As Moritz Hardt, Richard Zemel, and their colleagues argued in Fairness Through Awareness, machine learning models are highly efficient at identifying proxies. A zip code can act as a high-accuracy proxy for race; a browsing history can reconstruct gender; a list of hobbies can signal age.

When a model is trained to maximize utility, it will exploit any available signal to improve its predictions. If a sensitive attribute is correlated with the target variable, the model will inadvertently reconstruct that attribute from the remaining "neutral" features. This leads to indirect discrimination, where the model's decisions remain biased against protected groups despite the lack of explicit labels.

Detecting the Invisible Signal

How can we tell if a model has reconstructed a hidden attribute? A recurring mechanism in the literature is the use of an adversarial or "secondary" model.

In Lessons Learned from Problem Gambling Classification, researchers proposed a "model-matched" technique. To verify if a primary model is discriminating indirectly, they train a secondary model with the same architecture to predict the protected attribute (e.g., gender) using only the features available to the primary model. If this secondary model can predict the protected attribute better than random chance, it proves that the information is redundantly encoded in the features, and the primary model is likely utilizing it.

This concept is echoed in the work of Luca Deck and Maria De-Arteaga, who highlight in A Critical Survey on Fairness Benefits of XAI that redundant encoding analysis is a critical tool for auditing whether transparency at the feature level is actually sufficient to catch discriminatory patterns.

Repairing the Data and the Model

Once redundant encoding is detected, researchers have proposed two main ways to neutralize it: repairing the data or constraining the model.

1. Data Repair: Sorelle A. Friedler, Carlos Scheidegger, and Suresh Venkatasubramanian introduced a framework in Certifying and removing disparate impact to "repair" a dataset. Their algorithm transforms the marginal distributions of features for different groups so that they look identical to the model. By making the features statistically independent of the protected attribute, they physically remove the redundant encoding before the model ever sees the data.

2. Model Constraints: Alternatively, we can force the model to ignore these signals during training. Kazuto Fukuchi and Jun Sakuma proposed $\eta$-neutrality in Prediction with Model-Based Neutrality. This approach treats neutrality as a mathematical constraint within the Maximum Likelihood Estimation (MLE) process, effectively penalizing the model if its predictions can be used to reconstruct the sensitive attribute.

Formalizing the Search for Proxies

As the field matures, the focus has shifted from post-hoc detection to formal verification. Qusai Ramadan and Steffen Staab developed MBFair, a methodology that uses Linear Temporal Logic to identify hidden proxies during the software design phase. By mathematically modeling the flow of information, they can catch potential "attribute redundancy" before a single line of code is written.

While Benedikt Höltgen and Nuria Oliver have recently suggested in Reconsidering Fairness Through Unawareness from the Perspective of Model Multiplicity that FtU can sometimes work when multiple equally accurate models exist, the consensus remains: without explicitly addressing redundant encoding, "unawareness" is rarely a guarantee of fairness. True algorithmic equity requires an active awareness of how sensitive information leaks through the shadows of neutral data.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_mitigating_indirect_discrimination_via_r_45d09e3f,
  title = {Mitigating Indirect Discrimination via Redundant Encoding Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-indirect-discrimination-via-redundant-encoding-analysis/rdr_45d09e3f76a7480dbe8739a038c49f90},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_mitigating_indirect_discrimination_via_r_45d09e3f_v1,
  title = {Mitigating Indirect Discrimination via Redundant Encoding Analysis},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/mitigating-indirect-discrimination-via-redundant-encoding-analysis/rdr_45d09e3f76a7480dbe8739a038c49f90/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}