0.3.0 - 2023-08-14¶
βοΈ New Metrics Computation Mode¶
-
An
error_analysis
mode that measures subgroup and group metrics for correct and incorrect predictions, in addition to default groups. For example, when a default computation mode measures metrics for sex_priv and sex_dis, anerror_analysis
mode measures metrics for (sex_priv, sex_priv_correct, sex_priv_incorrect) and (sex_dis, sex_dis_correct, sex_dis_incorrect). Therefore, a user can analyze how a model is stable or certain about its incorrect predictions. -
An example yaml file for the computation mode:
dataset_name: COMPAS bootstrap_fraction: 0.8 n_estimators: 50 computation_mode: error_analysis sensitive_attributes_dct: {'sex': 1, 'race': 'African-American', 'sex&race': None}
π New Benchmark Fair-ML Datasets¶
-
LawSchoolDataset. A data loader for the Law School dataset that contains sensitive attributes among feature columns.
- Target: binary classify whether a candidate would pass the bar exam or predict a studentβs first-year average grade (FYA).
- Source: https://github.com/tailequy/fairness_dataset/blob/main/experiments/data/law_school_clean.csv.
- Broader description: https://arxiv.org/pdf/2110.00530.pdf.
-
RicciDataset. A data loader for the Ricci dataset that contains sensitive attributes among feature columns.
- Target: binary classify whether an individual obtains a promotion based on the exam results.
- Source: https://github.com/tailequy/fairness_dataset/blob/main/experiments/data/ricci_race.csv.
- Broader description: https://arxiv.org/pdf/2110.00530.pdf.
π Analyzers and Metrics¶
-
New subgroup metrics:
- Statistical Bias is a feature of a statistical technique or of its results whereby the expected value of the results differs from the true underlying quantitative parameter being estimated (ref).
- Aleatoric Uncertainty is a mean entropy of ensemble (ref).
- Overall Uncertainty is an entropy of mean prediction of ensemble (ref).
-
Changed a reference group in a sensitive_attributes_dct: now a disadvantaged group is used as a reference to compute intersectional metrics. For example, if we need to compute metrics for sex & race group (sex -- [male, female], race -- [white, black]), then sex&race_dis would include records for black females, and sex&race_priv would include all other records in a dataset.