A Normative approach to Attest Digital Discrimination

TL;DR

This paper introduces a normative framework for representing and checking digital discrimination in machine learning systems, aiming to make bias detection accessible to non-technical users.

Contribution

It formalizes non-discrimination norms using norms as an abstraction and proposes an algorithm to verify violations in ML systems.

Findings

Formalization of non-discrimination norms

Algorithm for checking norm violations in ML

Enhanced accessibility for bias detection tools

Abstract

Digital discrimination is a form of discrimination whereby users are automatically treated unfairly, unethically or just differently based on their personal data by a machine learning (ML) system. Examples of digital discrimination include low-income neighbourhood's targeted with high-interest loans or low credit scores, and women being undervalued by 21% in online marketing. Recently, different techniques and tools have been proposed to detect biases that may lead to digital discrimination. These tools often require technical expertise to be executed and for their results to be interpreted. To allow non-technical users to benefit from ML, simpler notions and concepts to represent and reason about digital discrimination are needed. In this paper, we use norms as an abstraction to represent different situations that may lead to digital discrimination. In particular, we formalise non-discrimination norms in the context of ML systems and propose an algorithm to check whether ML systems violate these norms.