Machine learning software can generate models that inappropriately discriminate against specific protected social groups (e.g., groups based on gender, ethnicity, etc.). Motivated by those results, software engineering researchers have proposed many methods for mitigating those discriminatory effects. While those methods are effective in mitigating bias, few of them can provide explanations on what is the root cause of bias.

Our goal is to detect better and mitigate algorithmic discrimination in machine learning software problems. Here we propose FairMask, a model-based extrapolation method that is capable of both mitigating bias and explaining the cause. In our FairMask approach, protected attributes are represented by models learned from the other independent variables (and these models offer extrapolations over the space between existing examples). We then use the extrapolation models to relabel protected attributes later seen in testing data or deployment time. Our approach aims to offset the biased predictions of the classification model by rebalancing the distribution of protected attributes.

The experiments of this paper show that, without compromising (original) model performance, FairMask can achieve significantly better group and individual fairness (as measured in different metrics) than benchmark methods. Moreover, compared to another instance-based rebalancing method, our model-based approach shows faster runtime and thus better scalability. When looking at individual fairness (as indicated by the Flip Rates), FairMask can ensure perfect individual fairness while other benchmarks cannot.

Based on the above, we conclude that: We can recommend FairMask for faster and more effective bias mitigation. FairMask greatly excludes the risk of individual unfairness: Two individuals who only differ in the protected attributes will always receive the same prediction outcomes.

Slide for this talk