Showing 1–1 of 1 results for author: Tamene, Y

Uncertainty Quantification for Local Model Explanations Without Model Access

Authors: Surin Ahn, Justin Grana, Yafet Tamene, Kristian Holsheimer

Abstract: We present a model-agnostic algorithm for generating post-hoc explanations and uncertainty intervals for a machine learning model when only a static sample of inputs and outputs from the model is available, rather than direct access to the model itself. This situation may arise when model evaluations are expensive; when privacy, security and bandwidth constraints are imposed; or when there is a ne… ▽ More We present a model-agnostic algorithm for generating post-hoc explanations and uncertainty intervals for a machine learning model when only a static sample of inputs and outputs from the model is available, rather than direct access to the model itself. This situation may arise when model evaluations are expensive; when privacy, security and bandwidth constraints are imposed; or when there is a need for real-time, on-device explanations. Our algorithm uses a bootstrap** approach to quantify the uncertainty that inevitably arises when generating explanations from a finite sample of model queries. Through a simulation study, we show that the uncertainty intervals generated by our algorithm exhibit a favorable trade-off between interval width and coverage probability compared to the naive confidence intervals from classical regression analysis as well as current Bayesian approaches for quantifying explanation uncertainty. We further demonstrate the capabilities of our method by applying it to black-box models, including a deep neural network, trained on three real-world datasets. △ Less

Submitted 24 June, 2023; v1 submitted 13 January, 2023; originally announced January 2023.