Toward a taxonomy of trust for probabilistic machine learning

Area of Science:

• * Computer Science, Artificial Intelligence, and Statistics.
• * Focus on probabilistic machine learning applications.

Background:

• * Probabilistic machine learning is crucial for decision-making in medicine, economics, and politics.
• * Establishing trust in these AI-driven decisions is paramount.
• * Existing research often focuses on specific trust-building aspects.

Purpose of the Study:

• * To develop a comprehensive taxonomy of trust breakdown points in probabilistic machine learning analyses.
• * To identify challenges and propose methods for enhancing trust across the entire analysis pipeline.
• * To highlight areas where trust-building is particularly difficult.

Main Methods:

• * Development of a four-step taxonomy for trust breakdown: goal translation, problem formulation, algorithm selection, and implementation.
• * Illustration of the taxonomy using two detailed case studies.
• * Review and description of methods to increase trust at each identified step.

Main Results:

• * Identified four critical stages where trust in probabilistic machine learning can be compromised.
• * Demonstrated how trust failures manifest at each stage through practical examples.
• * Cataloged diverse strategies for mitigating trust issues throughout the analytical process.

Conclusions:

• * The proposed taxonomy provides a structured approach to understanding and addressing trust in probabilistic machine learning.
• * It highlights specific challenges and opportunities for enhancing the reliability of AI systems.
• * Emphasizes the need for trust-building methods beyond current research concentrations.