What is Fairness in Machine Learning?
You'll learn why fairness is critical in ML systems and how to think about different types of fairness.
Why Fairness Matters
Machine learning models increasingly make consequential decisions about people's lives—loan approvals, hiring recommendations, medical diagnoses, criminal sentencing. When these models systematically disadvantage certain groups, they can perpetuate or amplify historical discrimination. Fairness in ML aims to ensure models treat individuals and groups equitably.
Two Core Fairness Paradigms
Statistical (Group) Fairness
This approach focuses on equal outcomes across demographic groups. For example, a loan approval model should have similar approval rates for equally qualified applicants regardless of race or gender. Think of it as "equal statistical treatment"—the model's predictions should have similar distributions across protected groups.
Individual Fairness
This perspective demands that similar individuals receive similar predictions. If two job applicants have nearly identical qualifications, the model should score them similarly—regardless of group membership. It's the "treat like cases alike" principle.
Context Is Everything
Here's the challenge: fairness isn't a universal mathematical property—it's deeply contextual. A "fair" credit scoring model might mean:
- Equal opportunity: qualified applicants have equal approval rates (emphasizes not missing deserving people)
- Demographic parity: all groups have equal approval rates (emphasizes equal outcomes)
- Equalized odds: both false positives and false negatives are balanced
These definitions can mathematically conflict! You often cannot satisfy all simultaneously. The right choice depends on legal requirements, stakeholder values, and the specific harm you're trying to prevent.
Key Takeaway: Fairness in ML requires balancing statistical group fairness with individual treatment, understanding that "fairness" is context-dependent and value-laden, not a single mathematical formula. Different applications demand different fairness definitions based on their potential harms.