Algorithmic Bias: Detection, Measurement, and Mitigation

Why Algorithmic Bias Matters

A machine-learning model trained on historical data can reproduce — and even amplify — patterns of discrimination that existed in that history. In high-stakes domains like loan approval, hiring, and healthcare, a biased model can cause real harm to real people. Regulators, employers, and professional standards increasingly expect data practitioners to detect and document bias before deployment. This lesson gives you the practical skills to do that.

Sources of Bias in a Machine-Learning Pipeline

Bias can enter a pipeline at several stages:

Historical bias exists in the real world before data collection begins. If lenders historically approved fewer loans for a particular demographic group, a model trained on those approvals will learn that pattern.

Representation bias occurs when a training dataset under-represents a group. If only 8% of training records belong to one demographic group, the model has too few examples to learn accurate decision boundaries for that group.

Measurement bias happens when features are recorded less accurately for some groups. For example, credit-bureau data may be thinner for recent immigrants, producing more missing values and less reliable scores.

Proxy features are variables that correlate strongly with a protected attribute even though they are not the attribute itself. ZIP code, for instance, reflects residential segregation and can serve as a proxy for race. Leaving such features in the model can reintroduce the very discrimination you tried to avoid by omitting race directly.

Label bias arises when the outcome label itself reflects past human judgment that was biased — for example, "default or not" labels that reflect discriminatory collection practices rather than true creditworthiness.

In your audit, you should document which stages produced the bias you find and explain the mechanism.

Three Fairness Metrics You Will Calculate

Fairness metrics compare model performance across demographic groups. There is no single metric that captures all dimensions of fairness; each answers a different question. You will use a threshold of 0.10: a difference above that threshold indicates a meaningful disparity.

Demographic Parity Difference (DPD) measures whether the model approves loans at the same rate for both groups regardless of their actual creditworthiness.

> DPD = P(Ŷ = 1 | Group A) − P(Ŷ = 1 | Group B)

A DPD of 0 means equal approval rates. A positive value means Group A is approved more often.

Equalized Odds Difference (EOD) checks whether both the true positive rate (TPR) and the false positive rate (FPR) are equal across groups. Equalized odds requires the model to be equally accurate for applicants who truly should be approved and equally cautious for those who should not.

> EOD = max(|TPR_A − TPR_B|, |FPR_A − FPR_B|)

Some implementations report TPR difference and FPR difference separately; consult the Fairlearn documentation for the exact form your version uses.

Predictive Parity (also called precision parity) asks: among all applicants the model approves, does the proportion who truly qualify differ by group?

> Predictive Parity Difference = Precision_A − Precision_B

where Precision = TP / (TP + FP) for each group.

These three metrics sometimes conflict with each other mathematically. Improving demographic parity can reduce equalized odds, and vice versa. This is not a bug in your calculations; it reflects a proven theoretical result. Document the trade-offs and justify your choice of mitigation.

Conducting a Bias Audit

1. Separate your confusion matrices. Filter the test dataset by demographic group and build a confusion matrix for each group independently. Never compute fairness metrics on the aggregate dataset — aggregation hides disparities.

2. Examine class imbalance. Use `value_counts()` on the label column within each group. Large imbalances mean the model saw far fewer positive examples for one group during training.

3. Inspect missing-value patterns by group. Use `df.groupby('group')['feature'].isna().mean()` to detect whether missingness is concentrated in one group. Imputation strategies that ignore group membership can propagate bias.

4. Identify proxy features. Compute correlation or mutual information between each feature and the protected attribute. Flag any feature with a notably high association.

5. Document everything in the provided bias-audit report template. Write a plain-language explanation of each finding suitable for a non-technical manager.

Mitigation Strategies

You will apply one pre-processing strategy in this lab:

• Sample reweighting assigns higher training weights to under-represented group-label combinations so the model treats them as more important. The `imbalanced-learn` library and Fairlearn both provide utilities for this.
• Proxy feature removal eliminates the feature most responsible for transmitting bias. After removal, retrain and recalculate metrics. Note that removal alone may not be enough if other correlated features remain.

After applying your chosen strategy, retrain the model and recalculate all three metrics. Present a before/after comparison table and write a paragraph interpreting the change. Acknowledge any metric that did not improve or that worsened.

Safety and Professional Responsibility

Data work does not involve physical hazards, but it carries ethical hazards with real consequences for people outside the room. Treat demographic data with care: use only the synthetic dataset provided; never upload real personal data to a shared server or public notebook. Follow your program's data-handling policy. When you write your audit report, be honest about limitations — a mitigation that improves one metric while degrading another must be disclosed, not hidden.