Interpretable Rule Discovery Through Bilevel Optimization of Split-Rules of Nonlinear Decision Trees for...

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Supervised classification demands both accuracy and interpretability for practical applications.
• Existing methods often struggle to balance classifier performance with human understanding.
• Interpretability is defined here as classifiers expressible in simple mathematical terms.

Purpose of the Study:

• To develop a novel approach for creating humanly interpretable classifiers.
• To represent classifiers as an assembly of simple mathematical rules using a nonlinear decision tree (NLDT).
• To ensure classifier interpretability by restricting split-rule structure and tree depth.

Main Methods:

• A nonlinear decision tree (NLDT) is proposed, where each node represents a nonlinear mathematical rule (split-rule).
• Split-rules partition data to minimize impurity in child nodes.
• An evolutionary bilevel optimization algorithm determines interpretable split-rules and their weights.

Main Results:

• The NLDT approach demonstrated encouraging performance on various test, benchmark, and industrial problems.
• The method was effective for problems with 2 to 500 features.
• The algorithm successfully balanced classifier accuracy with rule-based interpretability.

Conclusions:

• The proposed NLDT method offers a viable solution for achieving interpretable classification models.
• The approach shows potential for application to more complex and challenging classification tasks.
• This work contributes to the field of explainable artificial intelligence (XAI) in supervised learning.