A Differential-Evolution-Based Approach to Extract Univariate Decision Trees From Black-Box Models Using Tabular Data

The growing demand for complex machine learning models has increased the use of black-box models, such as random forests and artificial neural networks, posing significant challenges regarding explainability and interpretability. This manuscript addresses the critical problem of understanding and interpreting decisions from these opaque models, as a lack of interpretability can hinder their adoption in sensitive applications. To tackle this issue, we propose an evolutionary approach to induce univariate decision trees that accurately mimic the behavior of black-box models using tabular data. Our method employs two differential evolution algorithm variants, focusing on building univariate decision trees to enhance model explainability. Key contributions of this work include the development of a fitness function that balances accuracy with tree compactness to reduce overfitting and improve explanability. Additionally, we introduce a novel selection scheme that evaluates candidate solutions using synthetic instances, further enhancing the robustness against variance of the decision trees. Experimental results demonstrate that the proposed approach yields more precise and compact decision trees than traditional methods, significantly improving the explainability of complex machine learning models.