A comparative study of fairness-enhancing interventions in machine learning

Abstract

Computers are increasingly used to make decisions that have significant impact in people's lives. Often, these predictions can affect different population subgroups disproportionately. As a result, the issue of fairness has received much recent interest, and a number of fairness-enhanced classifiers and predictors have appeared in the literature. This paper seeks to study the following questions: how do these different techniques fundamentally compare to one another, and what accounts for the differences? Specifically, we seek to bring attention to many under-appreciated aspects of such fairness-enhancing interventions. Concretely, we present the results of an open benchmark we have developed that lets us compare a number of different algorithms under a variety of fairness measures, and a large number of existing datasets. We find that although different algorithms tend to prefer specific formulations of fairness preservations, many of these measures strongly correlate with one another. In addition, we find that fairness-preserving algorithms tend to be sensitive to fluctuations in dataset composition (simulated in our benchmark by varying training-test splits), indicating that fairness interventions might be more brittle than previously thought.

The Impact of Data Preparation on the Fairness of Software Systems

“…Although being aware of the drawbacks of fairness metrics like statistical parity, as these have been widely discussed in the literature [10], [23], we wanted to perform a not so common analysis of fairness that allowed us to compare the unfairness found in the predictions made by an ML model to that found in the data used to train that model. Nevertheless, our experiments have shown the brittleness of these metrics, as even those which were expected to show similar behaviours, such as CVS and DI, sometimes presented contradictory results [16]. This gives further confirmation that there is still room for improvement and progress when it comes to defining new fairness metrics.…”

“…This gives further confirmation that there is still room for improvement and progress when it comes to defining new fairness metrics. Specially in imbalanced scenarios, we believe that adopting more recently proposed fairness metrics based on group-conditioned performance [16] might be a small but crucial step towards achieving this goal. There is also room for improvement when it comes to the definition of individual fairness metrics, which seeks to treat similar individuals in a similar way [27].…”

“…Measures of this metric lie in [−1, 1], with 0 being optimal fairness. The sign of a measure indicates a skew in favour of either the protected or unprotected group [16]. Disparate impact (DI) is given by the ratio of the rate of favourable predictions for the protected group to that of the unprotected group [16]:…”

Designing Tools for Semi-Automated Detection of Machine Learning Biases: An Interview Study

“…Automatically detecting whether an ML model is biased against certain subgroups often involves segmenting a test set into subgroups and measures model performance on each to identify the subgroups on which the model underperforms [14]. A prerequisite for conducting this bias analysis is having the group membership of the input data.…”

mdfa: Multi-Differential Fairness Auditor for Black Box Classifiers

“…Similar to prior work on disparate impact [Feldman et al, 2015;Chouldechova, 2017] there is a need to re-balance the distribution of features conditioned on sensitive attributes. Our kernel matching technique deals with covariate shift [Gretton et al, 2009;Cortes et al, 2008], and has been used in domain adaptation (see e.g. [Mansour et al, 2009]) and counterfactual analysis (e.g.…”