Fig. 5: Explanations of the model’s loss rather than the model’s prediction yields new insights.
From: Explaining a series of models by propagating Shapley values
a We train on the first five cycles of NHANES (1999–2008) and test on the last three cycles (2009–2014). (b) We identify a simulated covariate shift in cycles 7–8 (2011-2014) by examining loss attributions. c Under a natural covariate shift, we identify and quantitatively validate test samples for which blood lead greatly increases the loss in comparison to training samples. d We ablate output attributions (G-DeepSHAP) and loss attributions (G-DeepSHAP, IME, KernelSHAP, and LIME) to show their respective impacts on model loss. We compare only to model-agnostic methods for loss attributions because explaining model loss requires explaining a series of models. Note that (b) and (c) show dependence plots (Supplementary Methods Section 1.3.2)).
