Group and Individual Fairness

Group vs. individual fairness

Many of us experience a phenomenon called collective responsibility. It is a situation when the inappropriate behavior of even a single member of a group causes punishment for the entire group. It is natural to consider it profoundly unjust. But wait, so far, we have discussed fairness in the context of groups. Does it mean we fall into this trap? The meaning of fairness for a group and an individual can differ. There is ongoing discussion if there is a trade-off between them or if they can be achieved simultaneously.

Group fairness

Group fairness is based on the assumption that bias is mainly caused by affiliation to a minority or historically discriminated subpopulation. This is often a valid case. However, there are some issues with this formulation. Firstly, we might not be aware of discrimination against specific groups. Even if we apply methods for mitigating issues with group fairness, it might turn out that there is a minority we haven’t considered. Therefore, the model is biased against them.

Some protected attributes can be perceived as a spectrum, not a bunch of buckets. For example, ethnic origin might not be unambiguous. Imagine a person whose parents are Latin and African-American—what is the correct value of this sensitive attribute?

There is one more issue: individuals can belong to multiple minority groups. In many approaches, it leads to many practical difficulties because we cannot easily analyze multiple subgroups (like gender and age). It can be even worse when only specific subgroup combinations are discriminated against (e.g., black females are discriminated against, but black males and white females are not). In such a case, we need to create auxiliary groups. It could be a young woman, an elderly woman, a young man, and so on. However, the number of created groups can be relatively large, and without a sufficient number of examples for each group. In such a case, we can risk heavy overfitting.