Fig. 6: Explaining a series of models comprised of a convolutional neural network feature extractor and a gradient boosted tree classifier.
From: Explaining a series of models by propagating Shapley values
a Explanations from G-DeepSHAP and state-of-the-art model-agnostic approaches. b Quantitative evaluation of approaches, including runtime and ablation of the top 10% of positive and negative features. Error bars are 95% confidence intervals based on 20 iterations of randomly drawing five explicand images, then computing attributions and ablation results.
