Interpret When Possible: A Tree-Based Hybrid Framework for Interpretable Classification

The demand for explainable artificial intelligence models continues to grow, yet achieving interpretable without compromising the predictive accuracy typical of complex black-box systems remains difficult. While existing hybrid approaches attempt to bridge this gap by combining white-box and black-box models, they often suffer from limitations, including reliance on simpler interpretable components (e.g., rules and linear models) and the inflexibility of requiring retraining to adjust transparency levels. To address these limitations, we propose a novel Tree-based Hybrid INterpretable (THIN) framework for classification. THIN integrates a compact classification tree with a high-performance black-box model, enabling dynamic control over transparency during inference through a decision-distance-based selection mechanism—without modifying the trained models. A tailored three-stage training algorithm is introduced to construct a high-quality decision tree under the guidance of the black-box model, enhancing its predictive performance while preserving interpretability. Experimental results on a benchmark and large-scale datasets demonstrate that THIN achieves predictive accuracy comparable to state-of-the-art black-box models, while maintaining significantly improved interpretability and lower model complexity compared to traditional decision trees.