FIG.07

Algorithmic Bias

Systematic and repeatable errors in a computer system that create unfair outcomes, such as privileging one arbitrary group of users over others.

Failure Domain
Demographic Disparity
Selection Bias
Historical Bias
Where Bias Enters

1. Data Collection

Risk: Under-representation, sampling bias, historical prejudice.

2. Labeling

Risk: Annotator bias, subjective ground truth definition.

3. Feature Selection

Risk: Proxy variables for protected attributes (e.g., zip code).

4. Objective Function

Risk: Optimizing for global accuracy ignores minority performance.

5. Deployment

Risk: Feedback loops, automation bias in human operators.
Fairness Metrics
Demographic Parity

Positive outcome rates are equal across subgroups, regardless of true label.

P(Ŷ=1 | A=0) = P(Ŷ=1 | A=1)
Equalized Odds

True positive and false positive rates are equal across subgroups.

P(Ŷ=1|Y=y, A=0) = P(Ŷ=1|Y=y, A=1)
Calibration

Predicted probabilities reflect true likelihood equally for all groups.

P(Y=1 | Ŷ=s, A=0) = P(Y=1 | Ŷ=s, A=1)
Subgroup AUC

Model discriminative power evaluated separately per demographic group.

AUC_grp1 vs AUC_grp2 DISPARITY
Subgroup Error Analysis
False Positive Rate (FPR) Comparison across Protected Attributes with 95% CI.
Key Insight
Group C experiences 3x higher false positive rate than the baseline, violating equalized odds.
Mitigation Strategies
1
Pre-Processing

Modify training data to remove underlying bias before training.

Reweighting Oversampling Data Augmentation
2
In-Processing

Modify learning algorithm to penalize discriminatory outcomes.

Adversarial Debiasing Fairness Constraints
3
Post-Processing

Adjust model predictions to enforce fairness constraints.

Threshold Optimization Calibration
Governance & Oversight
Model Cards & Datasheets

Standardized documentation of intended use, training data demographics, and known limitations.

Algorithmic Audits

Independent, third-party evaluation of system performance across intersectional subgroups.

Human-in-the-Loop

Mandatory operator review for high-stakes decisions with override authority.