A1075
Title: Spurious relationships can lead to algorithmic unfairness: Algorithmic bias explored using directed acyclic graphs
Authors: Soo Jung La - Seoul National University (Korea, South) [presenting]
Yongnam Kim - Seoul National University (Korea, South)
Abstract: Algorithmic fairness has become a central concern in the use of machine learning systems, particularly in high-stakes decision-making contexts. The aim is to use directed acyclic graphs (DAGs) as a formal tool to examine the causal structures underlying algorithmic decision processes. It is demonstrated that apparent algorithmic bias can emerge even when no true bias exists in the data-generating process, arising solely from the way algorithms exploit statistical association in data without regard for underlying causality. Specifically, causal structures are identified where biased predictions are artifacts of spurious pathways rather than evidence of actual discrimination. Building on this insight, common bias mitigation strategies pre-processing, in-processing, and post-processing are analyzed through the lens of DAGs to clarify their mechanisms and limitations. The framework highlights when and how these methods succeed or fail, depending on the structure of the causal graph. By grounding fairness interventions in causal reasoning, the aim is to offer a more principled understanding of algorithmic bias.
