Data can be generated by an unfair mechanism in numerous scenarios. For instance, a judicial system is unfair if it rejects the bail plea of an accused based on race, gender, or any other sensitive attributes. Similarly, the Stop, Question, and Frisk (SQF) dataset generated by the police who search for contraband based on the sensitive attributes of the pedestrian (Evans, 2017) is unfair. To mitigate the unfairness in the procedure generating the dataset, we need to not only identify the sources of unfairness, but also quantify the unfairness in these sources, quantify how these sources affect the overall unfairness towards a subset of sensitive attributes in a particular decision, and prioritize the sources before addressing the real-world issues underlying them.



Prior work of (Zhang, 2017) identifies and removes discrimination after data is generated but does not suggest a methodology to mitigate unfairness in the data generation phase. In our work, we assume a Causal Bayesian Network to model the data generating mechanism and use the notion of an unfair edge, same as (Chiappa, 2018) to be a source of discrimination and quantify unfairness along an unfair edge. We also bound the cumulative unfairness in a particular decision (e.g. Granting judicial bail) towards a subset of sensitive attributes (e.g. African Americans) in terms of edge unfairness in the unfair edges (e.g. Edge Race -> Judicial Bail Decision) and prove that there is no cumulative unfairness after eliminating edge unfairness in all the unfair edges.



Using the bounds of cumulative unfairness, we propose an alternative to the discrimination removal algorithm discussed in (Zhang, 2017) that gets rid of the path-specific effect constraints that grow exponentially in the number of sensitive attributes and values taken by them. We also design our algorithm in such a way that it is independent of the threshold of discrimination that is subjectively chosen unlike the algorithm designed in (Zhang, 2017). Finally, we discuss a priority algorithm for policymakers to address the real-world issues underlying the edges that result in unfairness. The experimental section validates the model assumption made to quantify edge unfairness.