Visual Exploration of Machine Learning Model Behavior with Hierarchical Surrogate Rule Sets

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Visualization

Background:

• Model interpretability is crucial for understanding complex machine learning models.
• Surrogate models, like decision trees and rule sets, are common interpretation methods.
• Existing methods face challenges with deep trees and large rule sets, hindering effective model explanation.

Purpose of the Study:

• To develop novel algorithmic and interactive solutions for improved model interpretation using surrogate models.
• To address the limitations of existing decision trees and rule sets for complex model approximation.
• To enhance the understandability and comparability of surrogate rules for tabular data.

Main Methods:

• Introduction of Hierarchical Surrogate Rules (HSR), an algorithm for generating hierarchical rules.
• Development of SuRE, a visual analytics system integrating HSR and a novel Feature-Aligned Tree visualization.
• Evaluation of HSR algorithm for parameter sensitivity, time performance, and comparison with surrogate decision trees.
• Usability and observational studies with domain experts to assess the effectiveness of the SuRE system and Feature-Aligned Tree visualization.

Main Results:

• The HSR algorithm demonstrates reasonable scalability and overcomes shortcomings of traditional surrogate decision trees.
• The Feature-Aligned Tree visualization effectively depicts rules as trees, aligning features for easier comparison.
• Domain experts using the SuRE system and Feature-Aligned Trees achieved very high accuracy on non-trivial interpretation tasks.
• The study provides insights into rule analysis task characterization for future visualization design.

Conclusions:

• Hierarchical Surrogate Rules (HSR) and the SuRE system offer a promising approach to enhance model interpretability for tabular data.
• The Feature-Aligned Tree visualization significantly improves the ability to perform complex rule analysis and comparisons.
• The developed methods and system are effective for domain experts in understanding and interacting with complex model behaviors.