Classification Metrics

import numpy as np

from langfair.metrics.classification import ClassificationMetrics

1. Introduction

Large language models (LLMs) used in classification use cases should be assessed for group fairness (if applicable). Similar to traditional person-level classification challenges in machine learning, these use cases present the risk of allocational harms. LangFair offers the following classification fairness metrics from the LLM fairness literature:

• Predicted Prevalence Rate Disparity (Feldman et al., 2015; Bellamy et al., 2018; Saleiro et al., 2019)

• False Negative Rate Disparity (Bellamy et al., 2018; Saleiro et al., 2019)

• False Omission Rate Disparity (Bellamy et al., 2018; Saleiro et al., 2019)

• False Positive Rate Disparity (Bellamy et al., 2018; Saleiro et al., 2019)

• False Discovery Rate Disparity (Bellamy et al., 2018; Saleiro et al., 2019)

2. Assessment

# Simulate dataset for this example. In practice, users should replace this data with predicted classes generated by the LLM,
# corresponding ground truth values, and corresponding protected attribute group data.
sample_size = 10000
groups = np.random.binomial(n=1, p=0.5, size=sample_size)
y_pred = np.random.binomial(n=1, p=0.3, size=sample_size)
y_true = np.random.binomial(n=1, p=0.3, size=sample_size)

Classification Metrics

---

ClassificationMetrics() - Pairwise classification fairness metrics (class)

Class parameters:

• metric_type - ({‘all’, ‘assistive’, ‘punitive’, ‘representation’}, default=’all’) Specifies which metrics to use.

Methods:

1. evaluate - Returns min, max, range, and standard deviation of metrics across protected attribute groups.

   Method Parameters:

   • groups - (array-like) Group indicators. Must contain exactly two unique values.

   • y_pred - (array-like) Binary model predictions. Positive and negative predictions must be 1 and 0, respectively.

   • y_true - (array-like) Binary labels (ground truth values). Positive and negative labels must be 1 and 0, respectively.

   • ratio - (boolean) Indicates whether to compute the metric as a difference or a ratio

   Returns:

   • Dictionary containing fairness metric values (Dictionary).

Generate an instance of class ClassificationMetrics using default metric_type='all', which includes “assistive”, “punitive”, and “representation” metrics.

cm = ClassificationMetrics(metric_type='all')

# Metrics expressed as ratios (target value of 1)
cm.evaluate(groups=groups, y_pred=y_pred, y_true=y_true, ratio=True)

{'FalseNegativeRateParity': 0.9683960547735326,
 'FalseOmissionRateParity': 0.9682772917805723,
 'FalsePositiveRateParity': 0.9832027144990514,
 'FalseDiscoveryRateParity': 0.9750294817188464,
 'PredictedPrevalenceRateParity': 1.010318056277584}

# Metrics expressed as differences (target value of 0)
cm.evaluate(groups=groups, y_pred=y_pred, y_true=y_true, ratio=False)

{'FalseNegativeRateParity': 0.022435421009698087,
 'FalseOmissionRateParity': 0.009568167034658404,
 'FalsePositiveRateParity': 0.005089653684952955,
 'FalseDiscoveryRateParity': 0.01776013575685509,
 'PredictedPrevalenceRateParity': 0.0030867911196673647}