Conditional Use Accuracy Difference

Overview

The difference between the positive predicted value and negative predicted value for both privileged and unprivileged groups

Calculation

(PPV for privileged group + PPV for unprivileged group) - (NPV for privileged group + NPV for unprivileged group)

Source

Usage

Manually

Python

# Calculate conditional use accuracy difference
difference = conditional_use_accuracy_difference(df, protected_attribute, privileged_group, labels, positive_label, y_true)

print("Conditional Use Accuracy Difference:", difference)

Using Fairness Object

difference = (fo.compute(conditional_use_accuracy_difference))

Results

Conditional Use Accuracy Difference: 0.6333333333234445

These results are obtained by using the input data given in the Create Example Data page under Getting Started

Interpretation

The Conditional Use Accuracy Difference metric quantifies the difference in accuracy between two groups or populations based on a specific condition or attribute. A positive value indicates that the accuracy of the classification model is higher for the positive predictive values compared to the negative predictive values, while a negative value indicates the opposite.