SCoPE: Sampling with Confidence via Predictive Information for Local Explanations

Explainable AI (XAI) seeks to make predictions interpretable, a goal reinforced by regulations such as GDPR and the DPDP Act . Due to the growing complexity of modern systems and the use of pre-trained models, post-hoc approaches are often preferred for their broader applicability and scalability. Popular post-hoc model-agnostic approaches explain complex model predictions by fitting local surrogate models using constraints or prior assumptions in their sampling strategies. Mehods like LIME, GLIME, and US-LIME treat perturbation generation as a passive sampling problem, without adapting to the surrogate's current state of knowledge and methods like UnRAVEL, BayesLIME that adapt sampling to model feedback lack a principled separation of uncertainty sources. While parameter-space information gain methods provide a principled criterion for reducing uncertainty in the surrogate coefficients, reliable local explanations ultimately depend on predictive consistency in the neighborhood of the target instance. To address these limitations, we introduce SCoPE, an uncertainty-aware framework for local post-hoc explanations. It is based on the Expected Predictive Information Gain (EPIG) objective, which measures how informative a queried perturbation is about predictions in the local explanation region. In pdf we provide mathematical formulation and empirical results that demonstrate that \texttt{SCoPE} improves explanation stability (Jaccard similarity) and faithfulness (comprehensiveness, sufficiency), sample efficiency and runtime over state of the art baselines spanning passive (LIME, Tilia, US-LIME), constraint driven (GLIME, UnRAvEL), and posterior variance based (BayesLIME) sampling. Results report that SCoPE attains the best comprehensiveness and Jaccard on all three datasets and the best sufficiency on Adult and MNIST, indicating that our method explanations that are both more faithful and more stable. We also show that the smallest budget at which SCoPE matches BayesLIME's CCM is 46-94% fewer queries while running 1.8-7 times faster depending on the datsaset. Together, these results show that targeting predictive information gain in the local neighborhood produces faithful, stable explanations at a fraction of the query budget and runtime of existing Bayesian baselines.