Individual Fairness

The foundational definition by Cynthia Dwork and colleagues states that a model is individually fair if, for any two individuals x and y, the distance between their predictions is bounded by the distance between their inputs:

• Mathematical constraint: D(f(x), f(y)) ≤ d(x, y)
• d(x, y): A task-specific metric defining similarity in the input space
• D(f(x), f(y)): A metric measuring divergence in the output or prediction space
• Lipschitz condition: The mapping must be Lipschitz-continuous with constant 1, ensuring predictions don't diverge arbitrarily for similar inputs

The central challenge of individual fairness is defining a meaningful distance metric d(x, y) that captures similarity for a specific task. This metric is not universal:

• Construct validity: The metric must reflect the ground truth of what makes two individuals genuinely similar in the decision context
• Fair metric learning: Techniques that learn a metric from data while explicitly accounting for protected attributes to avoid encoding bias
• Oracle-based approaches: Using human domain experts or consensus panels to define or validate similarity judgments
• Adversarial metric learning: Training a metric that is simultaneously predictive of outcomes and invariant to sensitive attributes

Individual fairness and group fairness are complementary but can conflict. Understanding their relationship is critical for audit design:

• Granularity difference: Individual fairness provides per-pair guarantees; group fairness provides aggregate statistical guarantees
• Incompatibility results: Research shows no non-trivial individual fairness constraint can simultaneously satisfy common group fairness definitions like equalized odds
• Metric amplification: A well-designed individual fairness metric can imply group fairness if the metric itself encodes group-level protections
• Hybrid frameworks: Modern approaches combine both paradigms, using individual fairness for local consistency and group fairness for global parity

Individual fairness is deeply related to causal and counterfactual fairness definitions, sharing a focus on the individual rather than the group:

• Causal distance metrics: Defining similarity using structural causal models, where two individuals are similar if they differ only in a protected attribute along non-causal paths
• Counterfactual pairs: Generating synthetic counterparts by intervening on protected attributes while preserving causally legitimate features
• Path-specific fairness: Decomposing the distance metric to isolate discriminatory causal pathways from legitimate ones
• Unified frameworks: Recent work unifies individual fairness with counterfactual fairness by defining the distance metric directly on the causal graph

Operationalizing individual fairness requires specialized auditing and enforcement methods beyond standard group-based tools:

• Lipschitz auditing: Empirically testing whether a model violates the Lipschitz condition by searching for pairs where D(f(x), f(y)) > d(x, y)
• Adversarial pair generation: Using gradient-based methods to find maximally violating pairs in the input space
• Regularization-based enforcement: Adding a Lipschitz penalty term to the training objective to enforce individual fairness during learning
• Post-processing correction: Applying a transformation to model outputs that enforces the distance constraint while minimizing prediction distortion

Despite its theoretical elegance, individual fairness faces significant practical hurdles that limit its standalone deployment:

• Metric specification burden: The requirement to define a complete distance metric for every task is often infeasible in complex, high-dimensional domains
• Computational cost: Pairwise auditing scales quadratically with dataset size, making exhaustive verification intractable for large-scale systems
• Subjectivity of similarity: Reasonable people can disagree on what constitutes a relevant similarity, introducing a new source of contestability
• Proxy vulnerability: A poorly specified metric can itself encode discriminatory assumptions, creating a false sense of fairness