Measuring Fairness of Text Classifiers via Prediction Sensitivity

TL;DR

This paper introduces a new fairness metric for text classifiers called accumulated prediction sensitivity, which measures how much predictions depend on protected attributes, aligning well with human perceptions and theoretical fairness notions.

Contribution

The paper proposes a novel prediction sensitivity-based fairness metric for text classifiers, linking it to group and individual fairness, and demonstrating its effectiveness through experiments.

Findings

The new metric correlates more strongly with human fairness judgments than existing metrics.

It has a theoretical basis connecting to statistical and individual fairness.

Experimental results on toxicity and bias datasets validate the metric's effectiveness.

Abstract

With the rapid growth in language processing applications, fairness has emerged as an important consideration in data-driven solutions. Although various fairness definitions have been explored in the recent literature, there is lack of consensus on which metrics most accurately reflect the fairness of a system. In this work, we propose a new formulation : ACCUMULATED PREDICTION SENSITIVITY, which measures fairness in machine learning models based on the model's prediction sensitivity to perturbations in input features. The metric attempts to quantify the extent to which a single prediction depends on a protected attribute, where the protected attribute encodes the membership status of an individual in a protected group. We show that the metric can be theoretically linked with a specific notion of group fairness (statistical parity) and individual fairness. It also correlates well with humans' perception of fairness. We conduct experiments on two text classification datasets : JIGSAW TOXICITY, and BIAS IN BIOS, and evaluate the correlations between metrics and manual annotations on whether the model produced a fair outcome. We observe that the proposed fairness metric based on prediction sensitivity is statistically significantly more correlated with human annotation than the existing counterfactual fairness metric.