Looking Deeper into Tabular LIME

In this paper, we present a thorough theoretical analysis of the default\nimplementation of LIME in the case of tabular data. We prove that in the large\nsample limit, the interpretable coefficients provided by Tabular LIME can be\ncomputed in an explicit way as a function of the algorithm parameters and some\nexpectation computations related to the black-box model. When the function to\nexplain has some nice algebraic structure (linear, multiplicative, or sparsely\ndepending on a subset of the coordinates), our analysis provides interesting\ninsights into the explanations provided by LIME. These can be applied to a\nrange of machine learning models including Gaussian kernels or CART random\nforests. As an example, for linear functions we show that LIME has the\ndesirable property to provide explanations that are proportional to the\ncoefficients of the function to explain and to ignore coordinates that are not\nused by the function to explain. For partition-based regressors, on the other\nside, we show that LIME produces undesired artifacts that may provide\nmisleading explanations.\n