AutoLIME and PWA-LIME: towards robust explanations of deep dynamical models

The increasing complexity of machine learning models highlights the need for interpretability, especially in critical domains requiring trust and transparency. Local Interpretable Model-agnostic Explanations (LIME) is a popular eXplainable AI (XAI) method that provides localized, instance-specific explanations using an interpretable surrogate model. However, its effectiveness is limited by the lack of systematic guidelines for tuning its hyperparameters. This paper addresses this limitation by proposing Automatic LIME (AutoLIME), a bi-level optimization framework to tune LIME’s kernel width. Additionally, we introduce PieceWise Affine LIME (PWA-LIME), a clustering-based extension of LIME for multi-instance explanations, particularly useful for interpreting black-box models of dynamical systems. Preliminary numerical results validate the potential of these methods in explaining opaque dynamical models.