ProbLog4Fairness: A Neurosymbolic Approach to Modeling and Mitigating Bias

Fairness definitions often conflict in practice: The paper highlights that multiple fairness definitions can be incompatible with each other, making it impractical to apply a one-size-fits-all approach to algorithmic fairness across different contexts.

Context-specific bias assumptions are more practical than universal fairness rules: Rather than trying to operationalize abstract fairness principles, the authors argue it's more effective to identify and model the specific biases relevant to each real-world task based on domain knowledge about systemic inequities in that context.

The approach successfully mitigated bias in both synthetic and real-world data: ProbLog4Fairness demonstrated effectiveness in reducing algorithmic bias across synthetic datasets with known biases as well as real-world tabular and image data, suggesting practical applicability beyond controlled settings.

Flexibility is a key advantage over existing methods: The framework outperforms baseline approaches because it can adapt to different types of bias rather than being locked into a single bias type or fairness notion, making it more suitable for diverse real-world applications.

Interpretability is built into the bias modeling process: By formalizing bias assumptions as logic programs, the approach maintains transparency in how biases are being identified and addressed, which is important for stakeholders to understand and trust bias mitigation efforts.