A unified framework for global and local interpretability using adaptive derivative-ordered random explanation
Experiments across three data modalities demonstrate that ADORE outperforms existing methods such as LIME and SHAP in handling complex interactions and computational efficiency, while providing detailed and reliable explanations.