ML interpretability: Simple isn't easy.

Area of Science:

• Artificial Intelligence
• Philosophy of Science
• Computer Science

Background:

• The importance of machine learning (ML) model interpretability is widely acknowledged.
• Current research often focuses on complex 'black-box' models like neural networks and methods for explainable AI (XAI).
• A clear definition and understanding of interpretability across different model types remain elusive.

Purpose of the Study:

• To clarify the fundamental nature of ML model interpretability.
• To explore the spectrum of interpretability, focusing on highly interpretable models.
• To analyze how varying degrees of interpretability are achieved in different ML models.

Main Methods:

• Examination of inherently interpretable models (linear models, decision trees).
• Analysis of models with partial interpretability (MARS, GAM).
• Philosophical and conceptual analysis of interpretability.

Main Results:

• Interpretability is not a monolithic concept; it varies across model types.
• Factors contributing to the interpretability of simpler models are identified.
• Methods for retaining interpretability in more complex models are explored.

Conclusions:

• While methods for achieving interpretability differ, its nature can be clearly defined for specific ML models.
• This work provides a clearer framework for understanding and evaluating ML interpretability.
• Further research can build on this clarified understanding for more transparent and trustworthy AI systems.