D-BIAS: A Causality-Based Human-in-the-Loop System for Tackling Algorithmic Bias

Human expertise is essential for bias detection: D-BIAS requires domain experts to identify unfair causal relationships in datasets, demonstrating that algorithmic bias mitigation cannot be fully automated—human judgment about what constitutes "fairness" in a specific context is a critical component of the solution.

Causal models provide transparency for bias auditing: The system uses graphical causal models to make the relationships between data features visible and interpretable, allowing users to understand why bias exists rather than just detecting that it does, which is crucial for accountability in high-stakes domains like hiring and healthcare.

Bias mitigation involves trade-offs that need human oversight: When users adjust causal relationships to reduce bias, D-BIAS shows the impact on fairness metrics, utility (model performance), data distortion, and data distribution—highlighting that removing bias often comes with costs that stakeholders must consciously evaluate and accept.

Fairness interventions should be context-specific: By enabling users to detect bias "against a group, say females, or a subgroup, say black females," the paper emphasizes that fairness solutions must account for intersectionality and different demographic contexts rather than applying one-size-fits-all debiasing approaches.

Debiased datasets enable fairer downstream applications: The tool produces modified datasets that users can deploy in real applications, suggesting that bias mitigation at the data preparation stage—rather than only at the model level—can be a practical pathway to more responsible AI systems.