Human resource management

A sample scenario in the context of human resource management illustrates the functioning of the Fairness Compass. The sensitive subgroups considered in this example are men and women. The question of interest is which definition of fairness would be most appropriate when it comes to assessing fairness in employee promotion decisions. Obviously, this is just a fictional thought experiment, and depending on the context, other answers with different results may apply. The purpose of the Fairness Compass is to support well informed decision making based on the defined requirements for a given scenario.

In the figure, the Fairness Compass is represented as a decision tree with three different types of nodes: The diamonds symbolize decision points; the white boxes stand for actions and the grey boxes with round corners are the fairness definitions. The arrows which connect the nodes represent the possible choices. After starting the process, the first question is about existing policies which may influence the decision. Fairness objectives can go beyond equal treatment of different groups or similar individuals. If the target is to bridge prevailing inequalities by boosting underprivileged groups, affirmative actions or quotas can be valid measures. Such a goal may stem from law, regulation, or internal organizational guidelines. This approach rules out any possible causality between the sensitive attribute and the outcome. If the data tells a different story in terms of varying base rates across the subgroups, this is a strong commitment which leads to subordinating the algorithm’s accuracy to the policy’s overarching goal. For example, many universities aim to improve diversity by accepting more students from disadvantaged backgrounds. Such admission policies acknowledge an equally high academic potential of students from sensitive subgroups and considers their possibly lower level of education rather as an injustice in society than as a personal shortcoming. 

For the sample scenario, the stakeholders may conclude that no such affirmative action policy is in place for promotion decisions. Therefore, they may choose “No” and document the reasoning behind their choice. This procedure is repeated question after question until a leaf node is reached which contains the recommended fairness definition. In this case, the outcome is “Equalized opportunities”, a concept which ensures that the probabilities of being correctly classified are the same for everyone.

Read full use case here.