Most discussions of algorithmic fairness focus on Group Fairness: ensuring that error rates or positive outcomes are balanced across broad demographic categories like race or gender. However, group fairness can be a blunt instrument. It is possible to satisfy group-level statistics while still being flagrantly unfair to specific people—a phenomenon known as "fairness gerrymandering."
Individual Fairness (IF) offers a more granular alternative. First proposed as the principle that "similar individuals should be treated similarly," it shifts the focus from aggregate populations to the person. As Shai Ben-David and colleagues argue in What’s Individual About Individual Fairness?, the defining characteristic of IF is that a violation must be provable using finite evidence. You don't need a census to know you were treated unfairly; you only need to point to a "peer" who is similar to you but received a better outcome.
The Metric is the Message
The core of individual fairness is the similarity metric. To say two people should be treated similarly, we must first define what "similar" means for a specific task. Mykola Pechenizkiy and Pratik Gajane point out in On formalizing fairness in prediction with machine learning that these metrics are not just mathematical tools; they are formalizations of "distributive justice."
If a bank decides that two loan applicants are "similar" if they have the same income and credit score, regardless of their zip code, the metric itself encodes a policy of geographic invariance. However, as noted across the literature—from Certification of Distributional Individual Fairness to Fairness Through Computationally-Bounded Awareness—the metric is the "Achilles' heel" of the field. If the metric is biased or ignores structural barriers, the resulting "fair" model will simply formalize those existing inequities.
Enforcing Fairness as Invariance
Once a metric is defined, how do we force a neural network to respect it? The research direction led by Mikhail Yurochkin and Yuekai Sun frames this as a problem of invariance.
In SenSeI: Sensitive Set Invariance for Enforcing Individual Fairness, they treat fairness like a safety constraint in robust optimization. Imagine a "sensitive subspace" in the data—a direction in the feature space that represents a protected attribute like gender. An individually fair model should be "smooth" or "invariant" along that direction. If you perturb an individual's data point along that sensitive axis, the model's prediction should not change.
This intuition is applied across different domains:
- Ranking: In Individually Fair Ranking, the authors use optimal transport to ensure that the relative ordering of individuals doesn't flip based on sensitive attributes.
- Optimization: SenSR (Sensitive Subspace Robustness) trains models to be robust against any changes within a predefined "fairness neighborhood."
- Post-processing: If you cannot retrain a model, Felix Petersen and colleagues suggest using Graph Laplacian regularization to "smooth" the predictions of an existing model, ensuring that nodes (people) connected by similarity edges receive similar scores.
From Heuristics to Formal Guarantees
As individual fairness methods matured, the focus shifted from "making models fairer" to "proving they are fair." This is the jump from empirical observation to certification.
In Individual Fairness Guarantees for Neural Networks, Matthew Wicker, Marta Kwiatkowska, and their team used Mixed-Integer Linear Programming (MILP) to provide global guarantees. They essentially "wrap" the neural network's non-linear functions in linear bounds, allowing them to mathematically prove that no two similar individuals in the entire input space will be treated differently.
Because MILP is computationally expensive, later work like Certification of Distributional Individual Fairness introduced "orthotopes"—convex shapes that approximate these fairness constraints—allowing researchers to certify models with thousands of neurons in minutes rather than hours.
Why It Works: The Bayesian and Max-Min Perspectives
The intuition for why these models become fairer often boils down to uncertainty and redistribution.
Alice Doherty and Andrea Patane found in Individual Fairness in Bayesian Neural Networks that Bayesian models are naturally fairer. By considering a distribution of possible weights rather than a single point estimate, Bayesian networks are less likely to "overfit" to biased patterns in the training data, leading to more consistent treatment of similar individuals.
In practical assignment problems, like peer review, Nihar B Shah and colleagues show in PeerReview4All that individual fairness can be achieved through max-min fairness. By prioritizing the "worst-off" individual (e.g., the paper with the lowest-quality reviewers), the system lifts the floor for everyone, proving that individual-level fairness constraints can actually improve the overall accuracy and integrity of the entire system.