Fairness metrics in ML Python - Model Metrics & Evaluation

Metrics & Evaluation - Fairness metrics

Which metric matters for fairness and WHY

Fairness metrics help us check if a model treats different groups equally. We want to avoid bias that can hurt some people unfairly. Common fairness metrics include Demographic Parity (checks if positive outcomes are equal across groups), Equal Opportunity (checks if true positive rates are equal), and Equalized Odds (checks if both true positive and false positive rates are equal). These metrics matter because they show if the model is fair in decisions, not just accurate.

Confusion matrix example for fairness

Imagine a model predicting loan approval for two groups: Group A and Group B.

    Group A Confusion Matrix:
      TP=40  FP=10
      FN=5   TN=45

    Group B Confusion Matrix:
      TP=30  FP=20
      FN=15  TN=35
    

We calculate True Positive Rate (Recall) for each group:

• Group A Recall = 40 / (40 + 5) = 0.89
• Group B Recall = 30 / (30 + 15) = 0.67

The difference shows Group B gets fewer correct positive predictions, indicating possible unfairness.

Tradeoff: Fairness vs Accuracy

Sometimes improving fairness can reduce overall accuracy. For example, if a model favors one group, it might be more accurate but unfair. Adjusting the model to treat groups equally might lower accuracy but is fairer. This tradeoff is important to understand and balance based on the problem.

Example: A hiring model might be very accurate but reject many qualified candidates from a minority group. Improving fairness means accepting more candidates from that group, which might slightly reduce accuracy but improves fairness.

What "good" vs "bad" fairness metrics look like

Good: Similar true positive rates and false positive rates across groups (e.g., TPR difference < 0.05). This means the model treats groups equally.

Bad: Large differences in metrics (e.g., one group has TPR 0.9, another 0.5). This means the model favors one group and is unfair.

Common pitfalls in fairness metrics

• Ignoring context: Fairness depends on the problem and groups; one metric does not fit all.
• Data imbalance: Small group sizes can make metrics unstable or misleading.
• Accuracy paradox: A model can be accurate but unfair if it ignores minority groups.
• Overfitting fairness: Adjusting too much for fairness on training data can hurt real-world performance.
• Ignoring multiple fairness aspects: Focusing on one metric may hide unfairness in others.

Self-check question

Your model has 95% accuracy overall but the true positive rate for a minority group is 40%, while for the majority group it is 85%. Is this model good for fairness? Why or why not?

Answer: No, this model is not fair. Even though accuracy is high, the minority group has a much lower true positive rate, meaning they get fewer correct positive predictions. This shows bias and unfair treatment.

Key Result

Fairness metrics measure if a model treats different groups equally by comparing rates like true positive rate across groups.

Practice

1. What is the main purpose of fairness metrics in machine learning models?

easy

A. To improve the speed of model training

B. To check if the model treats different groups equally

C. To increase the size of the training dataset

D. To reduce the number of model parameters

Solution

1. Step 1: Understand fairness metrics

   Fairness metrics are designed to measure if a model treats different groups fairly and without bias.

2. Step 2: Compare options

   Options B, C, and D relate to performance or data size, not fairness.

3. Final Answer:

   To check if the model treats different groups equally -> Option B

4. Quick Check:

   Fairness metrics = check equal treatment [OK]

Hint: Fairness metrics check equality across groups [OK]

Common Mistakes:

• Confusing fairness with accuracy
• Thinking fairness improves training speed
• Mixing fairness with data size

2. Which of the following is a correct way to calculate demographic parity difference between two groups?

easy

A. Sum of false negatives in both groups

B. Ratio of training samples in each group

C. Difference in positive prediction rates between groups

D. Difference in model training times

Solution

1. Step 1: Recall demographic parity definition

   Demographic parity measures if the positive prediction rates are similar across groups.

2. Step 2: Evaluate options

   Difference in positive prediction rates between groups matches this definition. Options A, C, and D are unrelated to demographic parity.

3. Final Answer:

   Difference in positive prediction rates between groups -> Option C

4. Quick Check:

   Demographic parity = positive rate difference [OK]

Hint: Demographic parity = compare positive rates [OK]

Common Mistakes:

• Using false negatives instead of positive rates
• Confusing sample ratios with fairness metrics
• Mixing training time with fairness

3. Given the following confusion matrices for two groups, what is the equal opportunity difference? Group A: TP=40, FN=10, FP=20, TN=30 Group B: TP=30, FN=20, FP=15, TN=35

medium

A. 0.25

B. 0.10

C. 0.05

D. 0.20

Solution

1. Step 1: Calculate true positive rates (TPR) for both groups

   TPR = TP / (TP + FN). For Group A: 40 / (40 + 10) = 0.8. For Group B: 30 / (30 + 20) = 0.6.

2. Step 2: Find equal opportunity difference

   Equal opportunity difference = |0.8 - 0.6| = 0.20.

3. Final Answer:

   0.20 -> Option D

4. Quick Check:

   Equal opportunity difference = |TPR_A - TPR_B| = 0.20 [OK]

Hint: Equal opportunity = difference in true positive rates [OK]

Common Mistakes:

• Confusing false positive rate with true positive rate
• Not dividing by (TP + FN)
• Ignoring absolute difference

4. You have this Python code to compute statistical parity difference but it gives wrong results:

def stat_parity_diff(y_true, y_pred, group):
    pos_rate_group1 = sum(y_pred[group == 1]) / len(y_pred[group == 1])
    pos_rate_group0 = sum(y_pred[group == 0]) / len(y_pred[group == 0])
    return pos_rate_group0 - pos_rate_group1

What is the main bug?

medium

A. The difference sign is reversed; should be group1 - group0

B. Using y_true instead of y_pred for positive rates

C. Not handling division by zero if group has no samples

D. Using sum instead of mean for positive rate

Solution

1. Step 1: Understand statistical parity difference

   It is usually calculated as positive rate of group 1 minus positive rate of group 0.

2. Step 2: Check code difference calculation

   The code returns pos_rate_group0 - pos_rate_group1, which reverses the sign.

3. Final Answer:

   The difference sign is reversed; should be group1 - group0 -> Option A

4. Quick Check:

   Sign order matters in parity difference [OK]

Hint: Stat parity = group1 rate minus group0 rate [OK]

Common Mistakes:

• Mixing up group order in subtraction
• Ignoring zero division errors
• Confusing y_true and y_pred

5. You want to ensure your loan approval model is fair by equalizing false positive rates (FPR) across two groups. Given these FPRs: Group X: 0.10 Group Y: 0.25 Which action best reduces this fairness gap?

hard

A. Adjust the model threshold for Group Y to be more strict

B. Increase training data size for Group X

C. Remove Group Y samples from training

D. Use accuracy as the only metric to evaluate fairness

Solution

1. Step 1: Understand false positive rate difference

   Group Y has a higher FPR (0.25) than Group X (0.10), meaning more false alarms for Group Y.

2. Step 2: Choose action to reduce FPR gap

   Adjusting the decision threshold for Group Y to be more strict reduces false positives, lowering FPR closer to Group X.

3. Final Answer:

   Adjust the model threshold for Group Y to be more strict -> Option A

4. Quick Check:

   Threshold tuning reduces FPR gap [OK]

Hint: Higher threshold for high FPR group to reduce false positives [OK]

Common Mistakes:

• Ignoring threshold adjustment
• Removing data harms fairness
• Using accuracy alone misses fairness issues