Did You Do Your Homework? Raising Awareness on Software Fairness and Discrimination

Abstract

Machine Learning is a vital part of various modern day decision making software. At the same time, it has shown to exhibit bias, which can cause an unjust treatment of individuals and population groups. One method to achieve fairness in machine learning software is to provide individuals with the same degree of benefit, regardless of sensitive attributes (e.g., students receive the same grade, independent of their sex or race). However, there can be other attributes that one might want to discriminate against (e.g., students with homework should receive higher grades). We will call such attributes anti-protected attributes. When reducing the bias of machine learning software, one risks the loss of discriminatory behaviour of anti-protected attributes. To combat this, we use grid search to show that machine learning software can be debiased (e.g., reduce gender bias) while also improving the ability to discriminate against anti-protected attributes.

Multi-objective search for gender-fair and semantically correct word embeddings

“…To improve the fairness-performance trade-off for ML models, Chen et al [45] used an ensemble approach, which combined models trained for different objectives (i.e., fairness and performance metrics). Hort and Sarro [46] showed that while bias of ML models can be reduced, this can come at the cost of losing the ability to differentiate between desired features.…”

Fairness Testing: A Comprehensive Survey and Analysis of Trends

“…Hort and Sarro [163] observed another side effect of fairness repair: it could cause loss of discriminatory behaviours of anti-protected attributes. Anti-protected attributes refer to the attributes that one might want the ML decision to depend upon (e.g., students with homework should receive higher grades).…”

Enhanced Fairness Testing via Generating Effective Initial Individual Discriminatory Instances

“…Fairness is a critical non-functional testing property of data-driven applications and machine learning software [42]. As such, it has received an increasing attention from both the software engineering [9,12,20,41] and machine learning research communities [6,37]. Among others, Brun et al [9] named this "software fairness" and called for software engineers to combat such discrimination and build fair software.…”