Course contentsShow
← Machine Learning and Deep Learning
Lesson 3289 of 3,53872. Bias and FairnessFree lesson

Individual Fairness: Treating Similar People Similarly

Define individual fairness through distance metrics: individuals at similar distances from the decision boundary should receive similar outcomes.

Individual Fairness: Treating Similar People Similarly

What you'll learn: How individual fairness uses distance metrics to ensure that similar individuals receive similar predictions, offering a fundamentally different fairness perspective than group-based notions.

The Core Idea

While group fairness metrics (like demographic parity or equalized odds) compare statistics across protected groups, individual fairness focuses on consistency at the person-to-person level. The principle: if two individuals are similar in all relevant respects, they should receive similar outcomes from your model.

The key challenge is defining "similar." Individual fairness relies on a distance metric that quantifies how alike two people are based on their features. Once you have this metric, the fairness constraint becomes: individuals at similar distances from the decision boundary should receive similar treatment.

Distance Metrics as the Foundation

Imagine two loan applicants with nearly identical credit scores, income, and debt ratios. If one is approved and the other rejected, individual fairness is violated—assuming those features capture all relevant similarity.

Mathematically, if d(x₁, x₂) measures the distance between individuals x₁ and x₂, and f(x) is your model's output, individual fairness requires:

Small d(x₁, x₂) → Small |f(x₁) - f(x₂)|

The model must be Lipschitz continuous with respect to your fairness metric: nearby inputs produce nearby outputs.

The Hard Part: Choosing the Right Metric

Defining "similar" requires domain expertise. Should age matter? Employment history? The metric encodes normative judgments about which differences are legitimate versus discriminatory. This is both individual fairness's strength (captures nuance) and weakness (requires careful, context-specific design).

Key Takeaway: Individual fairness enforces consistency by requiring similar predictions for individuals at similar distances in a carefully chosen feature space—complementing group fairness notions by protecting against arbitrary individual-level discrimination.