Explaining classifiers with measures of statistical association

Emanuele Borgonovo, Valentina Ghidini, Roman Hahn, Elmar Plischke

Abstract: A new class of probabilistic sensitivity measures that quantifies the degree of association between covariates and generic targets used in classification is proposed, and it is shown that such class possesses the zero-independence property. Corresponding estimators are introduced, asymptotic consistency is proven and bootstrap is used to quantify uncertainty in the estimates. The use of the new dependence measures as explanations in a statistical machine learning context is illustrated. The resulting approach, called Xi-method, is demonstrated through applications involving different data formats: tabular, visual and textual.