Mind the Graph When Balancing Data for Fairness or Robustness

TL;DR

This paper investigates how data balancing strategies impact fairness and robustness in machine learning, emphasizing the importance of understanding causal relationships to avoid unintended consequences.

Contribution

It provides conditions under which data balancing improves fairness and robustness, and highlights potential failure modes when causal structures are ignored.

Findings

Balanced distributions may not remove undesired dependencies as intended.

Ignoring causal graphs can lead to failure modes in data balancing.

Data balancing can interfere with other mitigation techniques like regularization.

Abstract

Failures of fairness or robustness in machine learning predictive settings can be due to undesired dependencies between covariates, outcomes and auxiliary factors of variation. A common strategy to mitigate these failures is data balancing, which attempts to remove those undesired dependencies. In this work, we define conditions on the training distribution for data balancing to lead to fair or robust models. Our results display that, in many cases, the balanced distribution does not correspond to selectively removing the undesired dependencies in a causal graph of the task, leading to multiple failure modes and even interference with other mitigation techniques such as regularization. Overall, our results highlight the importance of taking the causal graph into account before performing data balancing.