Fairness metrics in ML Python - Deep Dive

Overview - Fairness metrics

What is it?

Fairness metrics are ways to measure if a machine learning model treats different groups of people equally. They check if the model's predictions are biased or unfair towards certain groups based on attributes like race, gender, or age. These metrics help us understand and improve the fairness of AI systems. Without them, models might unintentionally harm or discriminate against some people.

Why it matters

Fairness metrics exist to prevent AI systems from making unfair decisions that can affect people's lives, such as in hiring, lending, or healthcare. Without fairness checks, biased models could reinforce social inequalities or cause harm. Using fairness metrics helps build trust in AI and ensures technology benefits everyone fairly.

Where it fits

Before learning fairness metrics, you should understand basic machine learning concepts like classification, prediction, and evaluation metrics such as accuracy and precision. After this, you can explore bias mitigation techniques and ethical AI practices to improve fairness in models.

Mental Model

Core Idea

Fairness metrics measure how equally a model treats different groups to detect and reduce bias in predictions.

Think of it like...

Imagine a teacher grading exams without looking at students' names or backgrounds to be fair. Fairness metrics are like rules that check if the teacher's grading is truly unbiased across all students.

┌───────────────────────────────┐
│          Model Output          │
├─────────────┬─────────────┬───┤
│ Group A     │ Group B     │...│
├─────────────┼─────────────┼───┤
│ Predictions │ Predictions │   │
│ (e.g., pass/fail)             │   │
├─────────────┴─────────────┴───┤
│ Fairness Metrics compare these│
│ predictions to check equality │
└───────────────────────────────┘

Build-Up - 7 Steps

1

Foundation

Understanding group fairness basics

Concept: Introduce the idea that fairness means treating groups equally in model predictions.

Fairness in machine learning often focuses on groups defined by sensitive attributes like gender or race. Group fairness means the model should perform similarly for these groups. For example, if a model predicts loan approvals, it should approve loans at similar rates for different groups if they have similar qualifications.

Result

Learners understand that fairness is about equal treatment across groups, not just overall accuracy.

Understanding group fairness lays the foundation for measuring bias and ensures we look beyond overall model performance.

2

Foundation

Basic classification metrics review

Concept: Review common metrics like accuracy, true positive rate, and false positive rate that fairness metrics build upon.

Accuracy measures how often the model is correct overall. True positive rate (TPR) is the chance the model correctly identifies positive cases. False positive rate (FPR) is the chance the model wrongly labels negatives as positives. These metrics help us see how well the model works for each group.

Result

Learners recall how to measure model performance, which is essential for fairness comparisons.

Knowing these metrics helps us compare model behavior between groups to spot unfair differences.

3

Intermediate

Demographic parity explained

🤔Before reading on: do you think demographic parity means equal positive rates or equal accuracy across groups? Commit to your answer.

Concept: Demographic parity requires that the model predicts positive outcomes equally often for all groups, regardless of true labels.

Demographic parity means the percentage of positive predictions should be the same for each group. For example, if 60% of Group A gets a positive prediction, then about 60% of Group B should too. This ignores whether predictions are correct or not.

Result

Learners see how demographic parity focuses on equal treatment in prediction rates, not accuracy.

Understanding demographic parity reveals how fairness can be about equal opportunity but may ignore actual qualifications.

4

Intermediate

Equalized odds and error rates

🤔Before reading on: do you think equalized odds requires equal true positive rates, false positive rates, or both across groups? Commit to your answer.

Concept: Equalized odds requires that both true positive rates and false positive rates are equal across groups.

Equalized odds means the model should be equally accurate for all groups when predicting positives and negatives. For example, the chance of correctly identifying a positive case (TPR) and wrongly labeling a negative case (FPR) should be the same for each group.

Result

Learners understand that equalized odds balances fairness with accuracy by equalizing error rates.

Knowing equalized odds helps balance fairness and correctness, avoiding unfair errors for any group.

5

Intermediate

Predictive parity and calibration

🤔Before reading on: does predictive parity mean equal positive prediction rates or equal accuracy among predicted positives? Commit to your answer.

Concept: Predictive parity means the accuracy among positive predictions is equal across groups.

Predictive parity requires that when the model predicts a positive, the chance that prediction is correct should be the same for all groups. For example, if 80% of positive predictions are correct for Group A, it should be about 80% for Group B too.

Result

Learners grasp how predictive parity focuses on fairness in the reliability of positive predictions.

Understanding predictive parity shows fairness can focus on trustworthiness of positive predictions, not just rates.

6

Advanced

Trade-offs between fairness metrics

🤔Before reading on: do you think all fairness metrics can be satisfied at the same time? Commit to your answer.

Concept: Different fairness metrics often conflict, making it impossible to satisfy all simultaneously.

In practice, satisfying demographic parity, equalized odds, and predictive parity at once is usually impossible, especially when groups have different base rates. Choosing which fairness metric to prioritize depends on the context and values of the application.

Result

Learners realize fairness is complex and requires careful choice of metrics based on goals.

Knowing fairness trade-offs prevents naive attempts to fix bias and encourages thoughtful metric selection.

7

Expert

Fairness metrics in real-world systems

🤔Before reading on: do you think fairness metrics alone guarantee fair AI in production? Commit to your answer.

Concept: Fairness metrics are tools but do not guarantee fairness without context, data quality, and ongoing monitoring.

In real systems, fairness metrics guide bias detection but must be combined with understanding data biases, stakeholder values, and legal requirements. Models need continuous fairness audits and adjustments as populations and contexts change.

Result

Learners appreciate that fairness is an ongoing process, not a one-time metric check.

Understanding real-world fairness challenges highlights the importance of combining metrics with human judgment and system design.

Under the Hood

Fairness metrics work by splitting data into groups based on sensitive attributes and calculating performance or prediction statistics separately for each group. They compare these statistics to detect disparities. Internally, this involves counting true positives, false positives, and other outcomes per group, then computing ratios or differences. These calculations reveal where the model treats groups differently.

Why designed this way?

Fairness metrics were designed to provide measurable, objective ways to detect bias in AI systems. Early AI models showed unintended discrimination, so researchers created these metrics to quantify fairness and guide improvements. Different metrics reflect different fairness philosophies and legal standards, acknowledging that fairness is complex and context-dependent.

┌───────────────┐       ┌───────────────┐
│   Dataset    │──────▶│ Split by Group │
└───────────────┘       └───────────────┘
         │                      │
         ▼                      ▼
┌───────────────┐       ┌───────────────┐
│ Group A Data │       │ Group B Data │
└───────────────┘       └───────────────┘
         │                      │
         ▼                      ▼
┌───────────────┐       ┌───────────────┐
│ Calculate    │       │ Calculate    │
│ Metrics (TPR,│       │ Metrics (TPR,│
│ FPR, etc.)   │       │ FPR, etc.)   │
└───────────────┘       └───────────────┘
         │                      │
         └──────────────┬───────┘
                        ▼
               ┌─────────────────┐
               │ Compare Metrics │
               │ for Fairness    │
               └─────────────────┘

Myth Busters - 4 Common Misconceptions

Quick: Does equal accuracy across groups guarantee fairness? Commit to yes or no.

Common Belief: If a model has equal accuracy for all groups, it must be fair.

Reality: Equal accuracy does not guarantee fairness because error types (false positives/negatives) can differ, causing unfair impacts.

Why it matters: Ignoring error types can lead to unfair harm, like one group facing more false accusations despite equal accuracy.

Quick: Can demographic parity be fair if groups have different real-world positive rates? Commit to yes or no.

Common Belief: Demographic parity always ensures fairness by equalizing positive prediction rates.

Reality: Demographic parity can be unfair if groups have different true positive rates, leading to ignoring real differences and possible harm.

Why it matters: Blindly enforcing demographic parity can cause qualified individuals in one group to be unfairly denied opportunities.

Quick: Is it possible to satisfy all fairness metrics at once? Commit to yes or no.

Common Belief: A model can satisfy all fairness metrics simultaneously.

Reality: Due to mathematical constraints, some fairness metrics conflict and cannot all be met at the same time.

Why it matters: Expecting all metrics to align can cause confusion and poor fairness decisions.

Quick: Does using fairness metrics alone guarantee an unbiased model? Commit to yes or no.

Common Belief: Applying fairness metrics ensures the model is unbiased and fair.

Reality: Fairness metrics detect bias but do not fix underlying data or societal biases; human judgment and context are needed.

Why it matters: Relying only on metrics can give a false sense of fairness and miss deeper issues.

Expert Zone

1

Fairness metrics depend heavily on the quality and representativeness of sensitive attribute data, which is often incomplete or noisy.

2

Choosing a fairness metric requires understanding the social and legal context, as different applications prioritize different fairness definitions.

3

Fairness metrics can be gamed or manipulated if used without transparency and accountability, leading to superficial fairness.

When NOT to use

Fairness metrics are less useful when sensitive attributes are unavailable or unreliable; in such cases, alternative approaches like causal fairness or individual fairness should be considered. Also, in highly dynamic environments, static fairness metrics may not capture evolving biases.

Production Patterns

In production, fairness metrics are integrated into model monitoring pipelines to detect drift in fairness over time. Teams use dashboards to track metrics per group and trigger alerts. Fairness audits combine metrics with qualitative reviews and stakeholder feedback to guide model updates.

Connections

Causal inference

Fairness metrics build on causal inference principles to understand if observed disparities are due to sensitive attributes or other factors.

Knowing causal inference helps distinguish correlation from causation in fairness, improving bias detection and mitigation.

Ethics in philosophy

Fairness metrics reflect ethical theories about justice and equality, connecting AI fairness to moral philosophy.

Understanding ethical foundations clarifies why different fairness definitions exist and how to choose among them.

Quality control in manufacturing

Both fairness metrics and quality control use statistical measures to detect deviations from standards across groups or batches.

Recognizing this similarity shows how fairness metrics apply general principles of monitoring and maintaining standards.

Common Pitfalls

#1 Ignoring differences in error types across groups.

Wrong approach:

Calculating only overall accuracy and assuming fairness:
accuracy = (TP + TN) / total
# No group-wise error analysis

Correct approach:

Calculate true positive rate and false positive rate per group:
TPR_group = TP_group / (TP_group + FN_group)
FPR_group = FP_group / (FP_group + TN_group)

Root cause:

Misunderstanding that equal accuracy alone does not capture unfair error distribution.

#2 Applying demographic parity without considering base rates.

Wrong approach:

Forcing equal positive prediction rates across groups regardless of actual positive rates:
if positive_rate_groupA != positive_rate_groupB:
adjust predictions to match

Correct approach:

Consider base rates and use metrics like equalized odds that account for true labels:
Ensure TPR and FPR are balanced rather than just positive rates.

Root cause:

Ignoring real differences in group distributions leads to unfair adjustments.

#3 Trying to satisfy all fairness metrics simultaneously.

Wrong approach:

Optimizing model to meet demographic parity, equalized odds, and predictive parity all at once without trade-offs.

Correct approach:

Select fairness metric aligned with application goals and optimize accordingly, acknowledging trade-offs.

Root cause:

Lack of awareness of mathematical incompatibilities between fairness definitions.

Key Takeaways

Fairness metrics help detect and measure bias by comparing model behavior across different groups.

No single fairness metric fits all situations; each captures different fairness ideas and trade-offs.

Understanding error types per group is crucial to assess fairness beyond overall accuracy.

Fairness is an ongoing process requiring metrics, context, and human judgment together.

Real-world fairness challenges include data quality, conflicting metrics, and evolving social norms.