Tuning Fairness by Balancing Target Labels

Abstract

The issue of fairness in machine learning models has recently attracted a lot of attention as ensuring it will ensure continued confidence of the general public in the deployment of machine learning systems. We focus on mitigating the harm incurred by a biased machine learning system that offers better outputs (e.g., loans, job interviews) for certain groups than for others. We show that bias in the output can naturally be controlled in probabilistic models by introducing a latent target output. This formulation has several advantages: first, it is a unified framework for several notions of group fairness such as Demographic Parity and Equality of Opportunity; second, it is expressed as a marginalization instead of a constrained problem; and third, it allows the encoding of our knowledge of what unbiased outputs should be. Practically, the second allows us to avoid unstable constrained optimization procedures and to reuse off-the-shelf toolboxes. The latter translates to the ability to control the level of fairness by directly varying fairness target rates. In contrast, existing approaches rely on intermediate, arguably unintuitive, control parameters such as covariance thresholds.

A Call to Action on Assessing and Mitigating Bias in Artificial Intelligence Applications for Mental Health

“…If bias is observed in the AI model as described above, then the researcher should return to the modeldevelopment stage and apply model in-processing or postprocessing bias-mitigation strategies (Berk et al, 2017;Gorrostieta et al, 2019;Kamishima et al, 2011;Woodworth et al, 2017;Zafar et al, 2017aZafar et al, , 2017b. For example, researchers could transform data, inject or recover noise (Calmon et al, 2017;Zhang et al, 2018), relabel the data to ensure an equal proportion of positive predictions for the sensitive group and its counterparts (Hardt et al, 2016;Luong et al, 2011), reweigh labels before training (Feldman et al, 2015;Kamiran & Calders, 2012;Luong et al, 2011), control target labels via a latent output (Kehrenberg et al, 2020), apply fairness regularization, penalize the mutual information between the sensitive feature and the classifier predictions (Kamishima et al, 2012), or add constraints to the loss functions that require satisfying a proxy for equalized odds or disparate impact (Woodworth et al, 2017;Zafar et al, 2017aZafar et al, , 2017b. See Figure 2 for a general framework for applying bias-mitigation techniques.…”

“…For biased labels, a researcher could relabel the data to ensure an equal proportion of positive predictions for the sensitive group and its counterparts (Hardt et al, 2016; Luong et al, 2011). Other techniques for improving problematic models include reweighing labels before training (Feldman et al, 2015; Kamiran & Calders, 2012; Luong et al, 2011) and controlling target labels via a latent output (Kehrenberg et al, 2020). It is critical that members of sensitive classes provide feedback on assigned labels, especially when those labels are subjective and culturally grounded (e.g., coding tweets for abusive language or labeling audio recordings for episodes of conflict).…”

An Empirical Comparison of Bias Reduction Methods on Real-World Problems in High-Stakes Policy Settings

“…Similarly, Iosifidis et al [34] used clustering across sensitive attribute and labels to come up with representative training data to train models, and Kamiran et al [37] explored multiple techniques involving sampling and re-weighing of training instances as pre-processing steps before applying machine learning models. Other popular preprocessing techniques involve relabelling and perturbation [39], details of which we omit from the paper.…”

A fuzzy-rough uncertainty measure to discover bias encoded explicitly or implicitly in features of structured pattern classification datasets

“…Moreover, they often require discretization of numeric sensitive features such as age, which can alter bias measures' outputs [7]. Individualbased fairness measures require strong assumptions such as the availability of an agreed-upon similarity metric, or knowledge of the underlying data generating process [13]. Additionally, they act as bias proxies as they do not measure bias directly.…”