Human Evaluation of Models Built for Interpretability.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Human-Computer Interaction

Background:

• Growing demand for interpretable machine learning (IML) systems.
• Lack of clear understanding regarding factors contributing to model interpretability.
• Decision sets as a specific class of logic-based models for IML.

Purpose of the Study:

• To advance the understanding of factors influencing human interpretability in decision set models.
• To identify specific types of complexity that significantly affect human performance in understanding models.
• To provide insights for designing more interpretable machine learning systems.

Main Methods:

• Controlled human-subject experiments conducted across two domains and three tasks.
• Focus on human-simulatability as a measure of interpretability.
• Analysis of complexity factors impacting user performance.

Main Results:

• Identified specific complexity factors that disproportionately affect human performance in understanding decision sets.
• Observed consistent trends in interpretability across different tasks and domains.
• Demonstrated that certain complexity types are more detrimental to interpretability than others.

Conclusions:

• Findings can guide the selection of regularizers in optimization to learn more interpretable models.
• Results suggest the existence of universal design principles for interpretable machine learning.
• Highlights the importance of human-simulatability in evaluating IML systems.