Protected attribute

As explained in Causal AI

A protected attribute is a characteristic, such as a demographic trait, whose use or effects in automated decisions are subject to fairness requirements.

A protected attribute is a personal characteristic, such as ethnicity or gender, whose use or influence in an automated decision raises fairness requirements. A loan model should not change its decision because a person belongs to a different ethnic group. section 8.1.3

Excluding the attribute from the input does not necessarily remove its influence. Neighborhood, social network, and shopping-habit features may act as proxies that let a model reproduce effects associated with ethnicity or another protected attribute. section 8.1.3

Counterfactual fairness analysis tests whether changing an attribute or its proxies would change the prediction:

• Identify features corresponding to the protected attribute, including possible proxies. section 8.1.3
• Ask whether the person would have received a different classification under the counterfactual attribute value. section 8.1.3
• Use the result to identify potential bias or inform training techniques intended to produce fairer algorithms. section 8.1.3

Loan approval

For a rejected applicant, ask: “Would this person have been approved if they were of a different ethnicity?” Evaluate corresponding changes in ethnicity proxies and check whether the outcome changes. section 8.1.3

Figure 8.3 For counterfactual fairness, suppose we want to test whether the algorithm has a bias against certain ethnicities. For a given feature vector, an ethnicity element, and a corresponding “reject” outcome, we test if the outcome would change if the ethnicity element changed.

Visualizes the counterfactual fairness test by comparing a rejection with the outcome after changing the ethnicity element. section 8.1.3

How Causal AI covers it

How deep it goes

Defines direct and proxy influence, then gives a counterfactual test; it stops short of solving broad AI fairness.

Examples

Uses loan approval, ethnicity, gender, neighborhood, social network, and shopping habits as concrete examples.

What it assumes

Assumes familiarity with machine-learning classifications and features while introducing proxies and counterfactual fairness.

What you can do after

Diagnose with counterfactual analysis whether changing a protected attribute or proxy could change an automated decision.

How it connects

Connects protected attributes to counterfactual fairness, algorithmic recourse, explainable AI, and proxy variables.

liveBookmanning.com

Read Causal AI on liveBook Get the book at manning.com

This explainer was generated by Manning's knowledge graph from the text of Causal AI. The section links point to the parts of the book it draws on.