Group Fairness vs Individual Fairness
What you'll learn: You'll understand the fundamental tension between enforcing equal outcomes across demographic groups and treating similar individuals similarly, regardless of their group membership.
Two Competing Fairness Paradigms
Imagine you're designing a loan approval system. Group fairness asks: "Do different demographic groups (defined by protected attributes like race or gender) receive approval at similar rates?" It focuses on statistical parity—ensuring that aggregate outcomes are balanced across groups.
Individual fairness asks instead: "Are two people who are similar in all relevant ways treated similarly?" This approach emphasizes that individuals shouldn't face discrimination based on irrelevant differences, focusing on case-by-case consistency rather than group statistics.
The Core Tension
These two approaches often conflict. Suppose your model learns that creditworthiness correlates with certain features that happen to differ across demographic groups. Enforcing group fairness (equal approval rates) might require treating some similar individuals differently just because they belong to different groups. Conversely, treating similar individuals identically might produce unequal group-level outcomes if the "similar" features themselves reflect historical inequities.
When to Use Each Approach
Group fairness is easier to measure and enforce—you can check approval rates, error rates, or prediction distributions across groups. It addresses systemic disparities and is often legally relevant.
Individual fairness is philosophically appealing and respects individual merit, but it's harder to implement—how do you define "similar" without encoding bias in your similarity metric? It also requires a valid distance function over individuals, which is challenging to construct fairly.
Key Takeaway: Group fairness optimizes for statistical parity across demographic groups, while individual fairness ensures similar treatment for similar individuals; these goals can conflict, and practitioners must make context-dependent tradeoffs between addressing systemic disparities and respecting individual similarity.