Classification Metrics

• Accuracy is the proportion of all classifications that were correct, whether positive or negative. \(( TP + TN) / all\)
• Recall, or true positive rate (TPR) is the proportion of all actual positives that were classified correctly as positives. \(TP / (TP + FN)\)
• False positive rate (FPR) is the proportion of all actual negatives that were classified incorrectly as positives (probability of false alarm). \(FP / (TN + FP)\)
• Precision is the proportion of all the positive classifications that are actually positive. \(TP / (TP + FP)\)

Google Developer’s ML Concepts

Fairness Metrics

Each metric requires that some aspects of the predictor behavior be comparable across groups.

A taxonomy of fairness metrics

More key metrics

Practice example

Two groups of 100 people.

Calculate accuracy, precision, true positive rate, and false positive rate

Practice example

Two groups of 100 people.

Same precision, similar true positive rate, but people in Group A who do not reoffend are 4x more likely to be wrongly flagged.

Group Discussion: Ethics frameworks

• Consequentialism: In the example, which error is worse: a false positive (wrongly flagged) or a false negative (missed)? Does your answer depend on the domain (bail, loans, medical care)?
• Deontology: Is it ever acceptable to use a protected attribute (e.g., race) to correct for bias?
• Contractualism: Which fairness definition could be justified to everyone affected, including the group with the higher base rate?
• Political economy: Who chooses the fairness metric? Who benefits from that choice?

Some references: Li et al. 2023, Trust Science 2019

The impossibility result

Kleinberg, Mullainathan & Raghavan (2016) and Chouldechova (2017) showed:

When base rates differ between groups, a (non-perfect) classifier cannot satisfy calibration/predictive parity and equal false positive and false negative rates at the same time.

• Fairness is not a bug to be fixed with better code
• Choosing a fairness definition is a value choice, not a technical one
• Someone must decide which errors matter most, and for whom

Fairness and the law

US anti-discrimination law (e.g., Title VII of the Civil Rights Act of 1964, Equal Employment Opportunity Act, Fair Housing Act, Equal Credit Opportunity Act) recognizes:

• Disparate treatment: intentionally treating people differently because of a protected characteristic
• Disparate impact: a neutral-looking practice that disproportionately harms a protected group (Griggs v. Duke Power Co., 1971)
• The EEOC’s four-fifths rule: a selection rate for one group below 80% of the rate for the highest group is evidence of adverse impact

Most algorithmic bias looks like disparate impact, not intent.

Laws generally only apply in employment, credit, and housing

Fairness and the law

Anti-discrimination protections broadly come in two forms: prohibitions on disparate treatment based on protected characteristics, and prohibitions on disparate impact based on those characteristics. Depending on the statute, protected characteristics can include race, color, national origin, religion, sex (including gender identity and sexual orientation), familial status, disability, age, and marital status.

From Bains 2024

Fairness in generative AI

• Image generators asked for “a CEO” or “a doctor” produce mostly white men; “a fast-food worker” produces mostly people with darker skin (Bloomberg analysis of Stable Diffusion, 2023)
• Attempts to correct this can overcorrect: in 2024, Google paused Gemini’s image generation of people after it produced historically inaccurate images (e.g., racially diverse 1940s German soldiers)
• Language models can associate names, dialects (e.g., African American English), and identities with negative traits

What would “fair” output even look like for a generative model?

References and Further Readings

• Barocas, S., Hardt, M., & Narayanan, A. (2023). Fairness and Machine Learning: Limitations and Opportunities. MIT Press.
• Chouldechova, A. (2017). Fair prediction with disparate impact. Big Data, 5(2), 153–163.
• Friedler, S. A., Scheidegger, C., & Venkatasubramanian, S. (2021). The (im) possibility of fairness: Different value systems require different mechanisms for fair decision making. Communications of the ACM, 64(4), 136-143.
• Kleinberg, J., Mullainathan, S., & Raghavan, M. (2016). Inherent trade-offs in the fair determination of risk scores.