Controlling Bias Between Categorical Attributes in Datasets: A Two-Step Optimization Algorithm Leveraging Structural Equation Modeling

Abstract

In the realm of data-driven systems, understanding and controlling biases in datasets emerges as a critical challenge. These biases, defined in this study as systematic discrepancies, have the potential to skew algorithmic outcomes and even compromise data privacy. Mutual information serves as a key tool in the analysis, discerning both direct and indirect relationships between variables. Utilizing structural equation modeling, this paper introduces a synthetic dataset generation method founded on a two-step optimization algorithm that aims to fine-tune variable relationships and achieve targeted mutual information levels between attribute pairs. The algorithm’s first phase utilizes gradient-less optimization, focusing on individual variables. The subsequent phase harnesses gradient-based methods to unravel deeper variable interdependencies. The approach is dual-purpose: it refines existing datasets for bias mitigation and creates synthetic datasets with defined bias levels, addressing a crucial research gap. Two case studies showcase the methodology. One emphasizes the finesse of network parameter adjustments in a simulated setting. The other applies the methodology to a realistic job hiring dataset, effectively reducing bias while safeguarding key variable relationships. In summary, this paper offers a novel method for bias management, presents tools for quantitative bias adjustments, and provides evidence of the method’s broad applicability through varied use cases.

Keywords

• Bias mitigation
• data fairness
• data generation
• explainable AI
• machine learning
• optimization