Information Theoretic Fair Representation Learning

99 concepts63 papers1 proposal

Machine learning models often inadvertently pick up on sensitive attributes like race, gender, or age because these traits are frequently correlated with target labels in historical datasets. Traditional fairness interventions typically try to fix this post-hoc by adding statistical constraints (such as demographic parity or equalized odds) directly to a model's final predictions. However, this approach can be brittle, task-specific, and easily bypassed if a downstream user applies a more complex or adversarial model to the outputs.

Instead of patching predictions downstream, a growing body of research focuses on sanitizing the data representations themselves. Information-theoretic fairness treats bias mitigation as a problem of controlling information leakage. By measuring the statistical dependence between a model's internal representations and protected attributes in bits, we can establish mathematical guarantees that no downstream model can exploit.


The Fair Information Bottleneck

To build a fair representation, we want to compress an input $X$ into a latent variable $Z$ that retains all information necessary to predict the target task $Y$, while completely purging any information about a sensitive attribute $S$.

This goal can be elegantly framed as a modified version of the classical Information Bottleneck (IB). In standard IB, a model maximizes the mutual information between the representation and the target, $I(Z; Y)$, while minimizing the information between the representation and the input, $I(Z; X)$, to encourage compression. In a fairness-constrained information bottleneck, as explored by Kim et al., the objective shifts: the model maximizes task utility $I(Z; Y)$ while explicitly minimizing mutual information with the protected attribute, $I(Z; S)$.

Input (X) ───► Latent Representation (Z) ───► Target Prediction (Y)
                     │
                     ▼ (Minimize Mutual Information)
             Sensitive Attribute (S)

By treating fairness as an information budget rather than a rigid penalty, frameworks like L-MIFR, developed by Song et al., allow practitioners to specify a maximum allowable level of information leakage in bits. This formulation enables the model to find the most expressive representation possible within a strict, quantifiable safety margin.


The Upper Bound Imperative

Why is minimizing mutual information $I(Z; S)$ difficult in practice? Mutual information is notoriously hard to compute exactly for continuous or high-dimensional neural network representations. Many early approaches relied on adversarial training (where a sub-network tries to guess $S$ from $Z$) or neural estimators like MINE to approximate it.

However, Gupta points out a fundamental flaw in using standard adversarial methods or lower-bound estimators for fairness: they optimize a lower bound of mutual information. If your goal is to eliminate bias, minimizing a lower bound provides no theoretical safety guarantees. It simply means your adversary or estimator is currently weak, leaving open the possibility that a more sophisticated downstream model could still extract the sensitive attribute from $Z$.

To solve this, modern frameworks focus on minimizing upper bounds of information leakage. For instance, Lowy et al. and Lowy introduce FERMI, which uses Exponential Rényi Mutual Information as a differentiable upper bound for multiple fairness criteria. Similarly, Kang et al. utilize a variational representation designed specifically for stable minimization, avoiding the training instability typically encountered when trying to decouple sensitive features.

When we successfully minimize a true upper bound on $I(Z; S)$, we get a provable guarantee: Luo et al. demonstrate that bounding this mutual information strictly limits demographic parity violations for any future downstream classification task.


Decomposing the Information Space

A common critique of fair representation learning is that it can create a zero-sum game between fairness and accuracy. If the sensitive attribute $S$ is naturally correlated with the target $Y$ in the real world, completely scrubbing $S$ from the representation will inevitably destroy some predictive signal. Jones formalizes this as the principle of representation scrubbing constraints, proving that removing sensitive information necessarily discards correlated utility in independent and identically distributed (IID) settings.

To mitigate this trade-off, researchers look at how information is shared across variables. Using Partial Information Decomposition (PID), Hamman et al. show that the mutual information between a sensitive attribute and a label is a fixed property of the dataset. Fairness interventions work by redistributing this information mass across different representational regions—shifting it away from unique sensitive components and toward redundant or shared components.

Instead of blindly deleting all shared information, frameworks like FD-VAE by Park et al. disentangle the latent space into three distinct views: target, protected, and mutual attributes. By re-encoding and aligning the mutual information subspace rather than discarding it entirely, they extract only the task-relevant components from the shared space. This alignment process preserves beneficial information for classification accuracy while filtering out the discriminatory bias inherent in the correlated data.


Generalizing Beyond Binary Classification

The information-theoretic paradigm easily generalizes beyond simple binary classification tasks:

  • Continuous Variables: When dealing with continuous sensitive attributes (like income or age), Ahuja leverages kernel machines to formulate convex formulations of KL divergence constraints, making optimization highly stable even for small datasets.
  • Intersectional Fairness: When auditing fairness across multiple overlapping groups (e.g., race, gender, and age simultaneously), the number of subgroups grows exponentially, leading to data sparsity. To handle this, the authors use total correlation—a multivariate extension of mutual information—to derive theoretical bounds on how much the dependency structure between protected attributes amplifies intersectional bias.
  • Post-Processing: If retraining a massive model from scratch is too expensive, Alghamdi et al. show that one can "tilt" the outputs of a pre-trained multi-class model using information projections, finding the closest fair probability distribution according to information-theoretic divergences without destroying the base classifier's utility.

Key Foundational Work

Learning Controllable Fair Representations (2018) — contributed 2 concept elements to this direction.
Fairness in Supervised Learning: An Information Theoretic Approach (2018) — contributed 2 concept elements to this direction.
Fairness in Supervised Learning: An Information Theoretic Approach (2018) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 70 papers
  2. Version 1 · 2026-10-04 · 119 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_information_theoretic_fairness_7baaaab1,
  title = {information-theoretic fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/information-theoretic-fairness/rdr_7baaaab15ef9458894f367a81ca0afe4},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 2, frozen text):

@misc{lacuna_information_theoretic_fairness_7baaaab1_v2,
  title = {information-theoretic fairness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/information-theoretic-fairness/rdr_7baaaab15ef9458894f367a81ca0afe4/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake