The goal of doing machine learning is frequently a prediction function; a tool that accepts a collection of available measurements and outputs a prediction for an unknown quantity of interest. Having obtained such a function, it is natural to ask "What is it doing?". There are a number of ways to answer this question and I will discuss techniques based on the functional ANOVA decomposition, which tries to represent such functions as sums of low-dimensional components. However, this representation is subject to distortion when the predictor variables are not independent and the prediction function extrapolates poorly. I propose a generalization of the functional ANOVA that corrects for this distortion at the cost of increased computation.
Groupe d’études et de recherche en analyse des décisions