B1608
Title: Statistics and explainability, an ideal partnership
Authors: Valentina Ghidini - Euler Institute (Switzerland) [presenting]
Abstract: Explainability is crucial, but some challenges hinder its applicability. For example, a formal definition of an explanation is usually missing, making it impossible to assess it in relation to theoretical or empirical benchmarks. Moreover, conventional methods do not typically offer insights into whether a model replicates the inherent dependency patterns of the process that generated the data. In an effort to address the above, employing statistical techniques is suggested to develop explanatory methods. To demonstrate the viability of this approach, a practical example of a newly defined method is provided. The explanations are defined as expected distances between probability distributions, which can be interpreted as variable importance measures. Notably, such explanations are backed by theoretical guarantees, as they are obtained through statistical estimators proven to be asymptotically consistent. The method provides pre hoc and post hoc explanations to compare estimated dependencies between the target and the covariates on both the data-generating process and the model of interest. Finally, such explanations can be obtained on any type of data coercible into a design matrix (tabular, image, text) and any regression or classification model. The method is also implemented in a Python package accessible on PyPI.
