Model exploration using conditional visualization

Abstract

Ideally, statistical parametric model fitting is followed by various summary tables which show predictor contributions, visualizations which assess model assumptions and goodness of fit, and test statistics which compare models. In contrast, modern machine‐learning fits are usually black box in nature, offer high‐performing predictions but suffer from an interpretability deficit. We examine how the paradigm of conditional visualization can be used to address this, specifically to explain predictor contributions, assess goodness of fit, and compare multiple, competing fits. We compare visualizations from techniques including trellis, condvis, visreg, lime, partial dependence, and ice plots. Our examples use random forest fits, but all techniques presented are model agnostic.