Machine learning models are notorious "cheaters." When tasked with predicting a target outcome $Y$ (like creditworthiness) from input data $X$, they will exploit any statistical correlation available. Often, this includes "sensitive attributes" $S$—such as race, gender, or age—that are embedded in the data through historical bias or proxy variables.
To prevent this, researchers have turned to Information Theory to provide a rigorous, mathematical language for fairness. By treating fairness as a problem of information flow, we can "scrub" discriminatory signals from data representations while preserving the utility of the model.
The Fairness Bottleneck
The core intuition behind this research direction is the creation of an intermediate, compressed representation $Z$. Instead of mapping $X$ directly to a prediction $\hat{Y}$, we map $X$ to a latent space $Z$ that acts as a filter.
As AmirEmad Ghassami, Sajad Khodadadian, and Negar Kiyavash describe in Fairness in Supervised Learning: An Information Theoretic Approach, this $Z$ serves as a bottleneck. The goal is to maximize the information $Z$ carries about the target $Y$ while minimizing the information it carries about the sensitive attribute $S$.
This is typically achieved through a multi-objective Lagrangian function: $$\min \mathcal{L} = \text{Error}(Y, \hat{Y}) + \lambda \cdot I(Z; S)$$ Here, $I(Z; S)$ is the Mutual Information (MI) between the representation and the protected attribute. By minimizing this term, we force the model to become "blind" to $S$.
Why Mutual Information?
Traditional fairness metrics often rely on simple linear correlations, which fail to capture complex, non-linear dependencies. Mutual information is a more powerful tool because it measures the total reduction in uncertainty about one variable given another.
Jiaming Song, Aditya Grover, and colleagues at Stanford developed the L-MIFR framework in Learning Controllable Fair Representations to give users direct control over this trade-off. They demonstrated that many disparate fairness algorithms are actually optimizing the same underlying information-theoretic objective, just with different penalty parameters.
Mapping Information to Fairness Criteria
Different definitions of "fairness" require different information-theoretic constraints:
- Statistical Parity: This requires the prediction $\hat{Y}$ to be independent of $S$. In information-theoretic terms, we minimize $I(\hat{Y}; S)$. Daniel McNamara and Cheng Soon Ong show in Provably Fair Representations that by bounding the conditional entropy of $S$ within $Z$, we can provide a theoretical upper limit on how much any downstream model can discriminate.
- Equalized Odds: This requires that for a given ground truth $Y$, the prediction $\hat{Y}$ is independent of $S$. This is enforced by minimizing the Conditional Mutual Information (CMI), $I(\hat{Y}; S | Y)$. Jianfeng Chi and Han Zhao apply this in Conditional Supervised Contrastive Learning for Fair Text Classification to ensure toxicity detectors perform equitably across demographic groups.
Scaling to Complex Scenarios
Recent work has moved beyond simple binary classification into more complex domains:
- Clustering: In Deep Fair Clustering via Maximizing and Minimizing Mutual Information, Pengxin Zeng and Peng Hu introduce FCMI. This framework simultaneously maximizes $I(X; \text{Clusters} | S)$ to ensure informative groupings while minimizing $I(\text{Clusters}; S)$ to ensure those groups aren't just proxies for protected attributes.
- Intersectional Fairness: Models often appear fair on individual attributes (race or gender) but fail on intersections (e.g., Black women). Narine Kokhlikyan and researchers at Meta AI address this "fairness gerrymandering" in Bias Mitigation Framework for Intersectional Subgroups in Neural Networks by minimizing the mutual information between the joint set of protected attributes and the model output.
- Generative Models: For tabular data generation, Tianchun Li and Tianci Liu developed UDF-DPO in Towards Universal Debiasing for Language Models-based Tabular Data Generation. They reformulate fairness as a preference learning task, where the model is fine-tuned to prefer data distributions with lower statistical dependencies between sensitive and target variables.
The Stability Advantage
A recurring theme across these papers is that direct MI optimization is often more stable than adversarial methods. As noted by Mattia Cerrato in Invariant Representations with Stochastically Quantized Neural Networks, information-theoretic penalties yield a more predictable, linear relationship between fairness constraints and performance compared to the "cat-and-mouse" instability of adversarial training. By treating fairness as a measurable quantity of information leakage, we move from heuristic "debiasing" to principled, provable equity.