There are several bias mitigators that can reduce algorithmic bias in machine learning models but, unfortunately, the effect of mitigators on fairness is often not stable when measured across different data splits. A popular approach to train more stable models is ensemble learning. Ensembles, such as bagging, boosting, voting, or stacking, have been successful at making predictive performance more stable. One might therefore ask whether we can combine the advantages of bias mitigators and ensembles? To explore this question, we first need bias mitigators and ensembles to work together. We built an open-source library enabling the modular composition of 10 mitigators, 4 ensembles, and their corresponding hyperparameters. Based on this library, we empirically explored the space of combinations on 13 datasets, including datasets commonly used in fairness literature plus datasets newly curated by our library. Furthermore, we distilled the results into a guidance diagram for practitioners. We hope this paper will contribute towards improving stability in bias mitigation.
INTRODUCTIONAlgorithmic bias and discrimination in machine learning are a huge problem. If learned estimators make biased predictions, they might discriminate against underprivileged groups in various domains including job hiring, healthcare, loan approvals, criminal justice, higher education, and even child care. These biased predictions can reduce diversity, for instance, in the workforce of a company or in the student population of an educational institution. Such lack of diversity can cause adverse business or educational outcomes. In addition, several of the above-mentioned domains are governed by laws and regulations that prohibit biased decisions. And finally, biased decisions can severely damage the reputation of the organization that makes them. Of course, bias in machine learning is a sociotechnical problem that cannot be solved with technical solutions alone. That said, to make tangible progress, this paper focuses on bias mitigators that can reduce bias in machine learning models. We acknowledge that bias mitigators can, at most, be a part of a larger solution.A bias mitigator either improves or replaces an existing machine learning estimator (e.g., a classifier) so it makes less biased predictions (e.g., class labels) as measured by a fairness metric (e.g., disparate impact). Unfortunately, bias mitigation often suffers from high volatility. There is usually less training data available for underrepresented groups.Less data means the learned estimator has fewer examples to generalize from for these groups.