A Review of Partial Information Decomposition in Algorithmic Fairness and Explainability.

Area of Science:

• Information Theory
• Machine Learning
• Algorithmic Fairness
• Explainability

Background:

• Partial Information Decomposition (PID) quantifies information shared among variables.
• Machine learning is increasingly used in high-stakes applications, necessitating fairness and explainability.
• Existing methods lack nuanced quantification of information contributions.

Purpose of the Study:

• To survey recent and emerging applications of PID in algorithmic fairness and explainability.
• To introduce a taxonomy of PID applications in these domains.
• To review PID estimation techniques and future directions.

Main Methods:

• Literature review of PID applications in algorithmic fairness and explainability.
• Taxonomic classification of PID roles: quantifying disparity, explaining contributions, and formalizing tradeoffs.
• Discussion of PID estimation techniques.

Main Results:

• PID enables disentangling non-exempt disparity in algorithmic fairness using causality.
• PID quantifies tradeoffs between local and global disparities in federated learning.
• A taxonomy highlights PID's utility in auditing, feature explanation, and federated learning.

Conclusions:

• PID offers powerful tools for enhancing algorithmic fairness and explainability.
• Further research is needed in PID estimation and application development.
• PID is crucial for responsible AI in high-stakes domains.