Risk mitigation in algorithmic accountability: The role of machine learning copies

Area of Science:

• Computer Science
• Artificial Intelligence
• Data Science

Background:

• Machine learning (ML) is a key driver of economic growth and efficiency.
• ML systems often integrate third-party components and APIs, increasing complexity.
• A significant challenge in ML is the lack of actionable accountability guidance.

Purpose of the Study:

• To investigate the role of copies in mitigating risks within complex ML systems.
• To explore the use of copies as an alternative for risk management when model retraining is not feasible.
• To provide actionable accountability guidance for ML systems.

Main Methods:

• Formally defining a copy as an approximated projection operator.
• Utilizing a conceptual framework of actionable accountability.
• Applying the approach to a real-world residential mortgage default dataset.

Main Results:

• Demonstrated the feasibility of using copies for risk mitigation in ML systems.
• Showcased the practical application of copies when model retraining or wrappers are not viable.
• Provided a viable alternative for enhancing accountability in ML.

Conclusions:

• Copies serve as a practical tool for risk mitigation in complex machine learning systems.
• The proposed method offers a viable solution for enhancing actionable accountability.
• The approach is effective even when direct model modification is not possible.