Dual feature-based and example-based explanation methods.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science

Background:

• Interpreting complex AI models remains a significant challenge.
• Existing explanation methods like LIME (Local Interpretable Model-agnostic Explanations) have limitations.
• Need for robust techniques for both local and global model understanding.

Purpose of the Study:

• Propose a new explainable AI (XAI) approach using convex hulls for local and global explanations.
• Develop a method for generating instance explanations via dual representations and convex combinations.
• Provide a computationally efficient alternative for feature importance calculation.

Main Methods:

• Constructing a convex hull around explained instances.
• Utilizing dual representations through convex combinations of polytope extreme points.
• Generating a dual dataset by sampling from the unit simplex.
• Training a dual linear surrogate model on the dual dataset.
• Computing feature importance via matrix calculations.

Main Results:

• The proposed method offers a novel way to achieve local and global model explanations.
• Dual representation facilitates example-based explanations.
• The approach is computationally efficient, relying on matrix calculations.
• Numerical experiments on real datasets validate the effectiveness of the approach.

Conclusions:

• The convex hull-based dual representation provides a powerful framework for explainable AI.
• This method enhances interpretability and offers an alternative to existing techniques.
• The approach is versatile and applicable across various domains requiring AI model transparency.