Exploring and Interacting with the Set of Good Sparse Generalized Additive Models.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Classical machine learning models offer limited interaction with domain experts.
• A single model output hinders collaborative refinement and selection.
• The Rashomon set, containing near-optimal models, provides a diverse solution space.

Purpose of the Study:

• To develop efficient algorithms for approximating the Rashomon set of sparse, generalized additive models.
• To enable domain experts to explore and select from a diverse set of near-optimal models.
• To address practical challenges in model interpretability and constraint satisfaction.

Main Methods:

• Algorithms for approximating the Rashomon set using ellipsoids for fixed support sets.
• Extension of methods to approximate Rashomon sets for various support sets.
• Experimental validation of the approximation fidelity and practical utility.

Main Results:

• Accurate approximation of the Rashomon set for sparse, generalized additive models.
• Demonstrated effectiveness in addressing variable importance, user-defined constraints, and shape function analysis.
• Validation of the approach's fidelity and practical problem-solving capabilities.

Conclusions:

• Approximating the Rashomon set facilitates crucial interaction between machine learning models and domain experts.
• The proposed methods provide a searchable space of diverse, near-optimal models.
• This approach offers practical solutions for model selection, constraint satisfaction, and interpretability.