How to Open a Black Box Classifier for Tabular Data

A lack of transparency in machine learning models can limit their application. We show that analysis variance (ANOVA) methods extract interpretable predictive from them. This is possible because ANOVA decompositions represent multivariate functions as sums fewer variables. Retaining the terms summation involving only one or two variables provides an efficient method to open black box classifiers. The proposed builds generalised additive (GAMs) by application L1 regularised logistic regression component retained decomposition logit function. resulting GAMs are derived using alternative measures, Dirac and Lebesgue. Both measures produce smooth consistent. term partial responses structured (PRiSM) describes family classifiers decompositions. demonstrate interpretability performance for multilayer perceptron, support vector machines gradient-boosting applied synthetic data several real-world sets, namely Pima Diabetes, German Credit Card, Statlog Shuttle UCI repository. shown be compliant with basic principles a formal framework interpretability.