metadata

title: Bias AUC
emoji: 🏆
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 3.19.1
app_file: app.py
pinned: false
license: apache-2.0

Bias AUC

Description of Metric

Suite of threshold-agnostic metrics that provide a nuanced view of this unintended bias, by considering the various ways that a classifier’s score distribution can vary across designated groups.

The following are computed where $D^{-}$ is the negative examples in the background set, $D^{+}$ is the positive examples in the background set, $D_{g}^{-}$ is the negative examples in the identity subgroup, and $D_{g}^{+}$ is the positive examples in the identity subgroup:

$\begin{aligned} \text{Subgroup AUC} &= \text{AUC} (D^{-}_{g} + D^{+}_{g} ) &(1)\\ \text{BPSN AUC} &= \text{AUC} (D^{+} + D^{-}_{g} ) &(2)\\ \text{BNSP AUC} &= \text{AUC} (D^{-} + D^{+}_{g} ) &(3) \end{aligned}$

How to Use

from evaluate import load

target = [['Islam'],
 ['Sexuality'],
 ['Sexuality'],
 ['Islam']]

label = [0, 0, 1, 1]

output = [[0.44452348351478577, 0.5554765462875366],
 [0.4341845214366913, 0.5658154487609863],
 [0.400595098733902, 0.5994048714637756],
 [0.3840397894382477, 0.6159601807594299]]

metric = load('Intel/bias_auc')

metric.add_batch(target=target,
                 label=label,
                 output=output)

metric.compute(subgroups = None)