Definition
To explain : sample around , evaluate , bin the features (tabular data) or switch superpixels on and off (images), fit a simple linear model locally and report its most prominent coordinates (Ribeiro, Singh, Guestrin 2016).
Formula
Local surrogate: for samples near ; explanation = the few with the largest .
Model-agnostic: it only needs to evaluate . Because it samples off the data manifold it can be fooled (Slack et al. 2020), and on the same decision it can disagree completely with SHAP or counterfactual methods.