CO-LIME: Correlation-aware Locally Interpretable Model-agnostic Explanations

As artificial intelligence (AI) systems are increasingly deployed in high-stakes domains such as healthcare and finance, it becomes extremely important to study the interpretability of them. Existing explanation methods, such as Local Interpretable Model-agnostic Explanations (LIME), rely on random perturbation mechanisms that assume feature independence to approximate local decision boundaries. However, these approaches often neglect the underlying joint distribution of the original data during perturbation generation, resulting in synthetic samples that deviate from the true data manifold and disrupt intrinsic feature dependencies. To address these limitations, we propose Correlation-aware Locally Interpretable Model-agnostic Explanations (CO-LIME), which employs Conditional Tabular Generative Adversarial Networks (CTGANs) to explicitly capture nonlinear dependencies through adversarial training. We also provide four methods to quantitatively assess the effectiveness of explanation methods by R² scores, mean absolute error (MAE), mean squared error (MSE), and mean median error (MedAE). Experimental results demonstrate that CO-LIME outperforms baseline methods, achieving a 55.05% improvement in R² and a 46.29% reduction in MSE.