Multi-Objective Evolutionary Rule-Based Classification with Categorical Data.

Area of Science:

• Machine Learning
• Data Mining
• Artificial Intelligence

Background:

• Model interpretability is crucial for validation and ethical evaluation.
• Existing methods may produce complex classifiers that are difficult to understand.
• Categorical data classification requires interpretable models.

Purpose of the Study:

• To propose a novel method for generating easily interpretable rule-based classifiers from categorical data.
• To optimize both classifier performance and rule set size.
• To enhance the ethical and legal evaluation of classification models.

Main Methods:

• A multi-objective optimization approach was employed.
• Two objectives were prioritized: maximizing classifier performance and minimizing the number of rules.
• Multi-objective evolutionary algorithms (ENORA and NSGA-II) were adapted to optimize performance using accuracy, ROC AUC, and RMSE metrics.

Main Results:

• The proposed method generated highly accurate and interpretable classification models.
• Extensive comparisons with classical methods (PART, JRip, OneR, ZeroR) were performed.
• Experiments utilized multiple validation modes (full training, 10-fold cross-validation, train/test split) on public datasets.

Conclusions:

• The novel method successfully generates classification models that are both highly accurate and easy to interpret.
• The approach offers a valuable tool for tasks requiring transparent and understandable predictive models.
• The findings support the use of multi-objective optimization for creating interpretable machine learning classifiers.