StrategyAtlas: Strategy Analysis for Machine Learning Interpretability.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Businesses hesitate to adopt complex machine learning (ML) models due to interpretability challenges.
• Existing ML explanation methods offer only local, instance-level insights, failing to capture global model behavior.

Purpose of the Study:

• To introduce a novel approach for understanding the global behavior of complex ML models using strategy clusters.
• To present StrategyAtlas, a system for analyzing, explaining, and utilizing these strategy clusters.
• To demonstrate the practical application of strategy clusters in improving ML models within a business context.

Main Methods:

• Identified strategy clusters as groups of data instances treated distinctly by ML models.
• Developed StrategyAtlas for the analysis and explanation of these model strategies.
• Applied the system in a real-world use case with an insurance company for automatic acceptance.

Main Results:

• Strategy clusters effectively reveal the global behavior of complex ML models.
• StrategyAtlas facilitates the exploration and understanding of these clusters for data scientists.
• Insights from strategy clusters enabled improvements to the production ML model.

Conclusions:

• Strategy clusters offer a powerful mechanism for global ML model interpretability.
• StrategyAtlas enhances data scientists' ability to understand and refine complex ML models.
• This approach bridges the gap between complex ML adoption and business needs in high-risk environments.