Definitions, methods, and applications in interpretable machine learning.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Machine learning models excel at prediction but interpreting their learned patterns is complex.
• Existing interpretation methods lack a unified framework, causing confusion.
• A clear understanding of interpretability is crucial for trustworthy AI.

Purpose of the Study:

• To define interpretability in machine learning.
• To introduce the predictive, descriptive, relevant (PDR) framework for evaluating interpretations.
• To categorize existing interpretation methods and suggest future research directions.

Main Methods:

• Defining interpretability and introducing the PDR framework.
• Categorizing interpretation methods into model-based and post hoc techniques.
• Illustrating the PDR framework with real-world examples.

Main Results:

• The PDR framework provides three key evaluation criteria: predictive accuracy, descriptive accuracy, and relevancy.
• Interpretation methods are categorized into model-based and post hoc, with subgroups like sparsity, modularity, and simulatability.
• The framework highlights the importance of human audience in judging relevancy.

Conclusions:

• The PDR framework offers a common vocabulary for discussing and choosing machine learning interpretation methods.
• This work aims to reduce confusion and facilitate better understanding of model interpretability.
• Future work should address limitations of current methods based on the PDR framework.