UCI Adult (Version 2) - Artificial Prevalence Protocol (APP) Test Indices

Dataset
Adult

# data points
45,222

# non-sensitive features (preprocessed)
95

sensitive attribute
sex

S = 1
Male

Pr(S=1)
0.675

target variable
income/class

Y = ⊕
>$50,000

Pr(Y=⊕)
0.248

The UCI Adult dataset is loaded using fetch_openml from scikit-learn (version 2). To ensure reproducibility, the dataset is split into three equal subsets using stratified sampling based on the joint distribution of (S, Y) and a random seed of 0.

The indices stored in adult_D1.indices, adult_D2.indices, and adult_D3.indices refer to the original dataset's instances after any instances containing missing values have been removed. The indices in adult_D3_protocol.indices correspond to the instances in the LabelledCollection constructed using the indices from adult_D3.indices.

Categorical features are converted to dummy variables, and numerical features are standardised using a standard scaler.

Feature Type
Features

Categorical
workclass, education, marital-status, occupation, race, native-country

Numerical
age, capital-gain, capital-loss, hours-per-week

The Artificial Prevalence Protocol (APP) is applied to vary the distribution of the sensitive attribute across 11 prevalence values: Pr(s|⊖) ∈ {0.0, 0.1, ..., 0.9, 1.0}. To minimise variance in the evaluation results, 500 samples are drawn repeatedly under the protocol. This entire process is repeated 10 times, consistently using a random seed of 0 for reproducibility.

This work has been funded by the QuaDaSh project “Finanziato dall’Unione europea- Next Generation EU, Missione 4 Componente 2 CUP B53D23026250001”