ML Ethics & Fairness

Types of Bias

1. Data Bias

Training data doesn't represent the real population.

• Example: Loan approval model trained only on white applicants → Biased against minorities
• Fix: Ensure balanced representation in training data

2. Algorithmic Bias

The algorithm itself is biased.

• Example: Using ZIP code as feature (proxy for race) → Redlining
• Fix: Remove protected attributes (race, gender, age)

3. Historical Bias

If the past was unfair, the model learns unfairness.

• Example: Hiring model trained on previous unfair hiring decisions
• Fix: Clean the labels or retrain with fair labels

Fairness Definitions

Statistical Parity

Equal positive prediction rate across groups.

P(predict=1 | Group A) = P(predict=1 | Group B)

Equalized Odds

Equal True Positive Rate and False Positive Rate.

• Everyone has equal chance of being caught (TP rate)
• Everyone has equal false alarm rate (FP rate)

Demographic Parity

Prediction should be independent of protected attribute.

Detecting Bias

from sklearn.metrics import confusion_matrix

# Check metrics by group
for group in ['Male', 'Female']:
    y_true_group = y_true[gender == group]
    y_pred_group = y_pred[gender == group]
    
    acc = accuracy_score(y_true_group, y_pred_group)
    # If accuracy_male >> accuracy_female, bias detected!

Mitigating Bias

Pre-processing

Clean training data before training:

• Remove protected attributes
• Balance representation
• Audit for proxy variables

In-processing

Modify algorithm during training:

• Fairness constraints in objective function
• Different regularization by group

Post-processing

Adjust predictions after training:

• Change threshold by group
• Example: If females have lower approval rate, lower threshold for females

Real-World Consequences

COMPAS Recidivism Score

Algorithm used to predict prisoner reoffending showed bias against Black Americans.

• False positive rate for Black defendants: 45%
• False positive rate for white defendants: 23%

Amazon Hiring ML

Trained on historical data where males were hired more.

• Model learned to prefer male applicants
• System scrapped after discovery

Loan Approval

Denying loans based on ZIP code (proxy for race) = Illegal (Redlining)