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

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• AI algorithms can learn and perpetuate social biases from training data.
• Deployment in critical domains like hiring and healthcare raises fairness and accountability concerns.
• Existing automated bias mitigation methods may lack transparency and user control.

Purpose of the Study:

• To introduce D-BIAS, a visual, interactive tool for auditing and mitigating social biases in tabular datasets.
• To enable users to detect and address bias through a human-in-the-loop approach using causal models.
• To provide a method for generating debiased datasets with minimal distortion and utility loss.

Main Methods:

• Utilizes a graphical causal model to represent feature relationships and incorporate domain knowledge.
• Employs an interactive interface for users to identify unfair causal relationships and apply bias mitigation strategies.
• Implements a novel simulation method to generate debiased datasets based on user interactions with the causal model.

Main Results:

• D-BIAS significantly reduces bias compared to baseline methods across various fairness metrics.
• The tool achieves bias reduction with minimal data distortion and a small loss in utility.
• User studies indicate that the human-in-the-loop approach enhances trust, interpretability, and accountability over automated methods.

Conclusions:

• D-BIAS offers an effective and transparent approach to auditing and mitigating social biases in AI.
• The visual, interactive, and human-in-the-loop nature of D-BIAS improves fairness, trust, and interpretability in machine learning applications.
• D-BIAS empowers users to create fairer datasets for downstream applications, promoting responsible AI deployment.