In plain words: Splitting credit for a prediction among input features (Shapley values) hinges on how you define an "absent" feature, and small changes there swing the numbers. The unified setup adds confidence intervals and reveals a popular choice credits features the model never uses.
Abstract · The Explanation Game: Explaining Machine Learning Models Using Shapley Values
A number of techniques have been proposed to explain a machine learning model's prediction by attributing it to the corresponding input features. Popular among these are techniques that apply the Shapley value method from cooperative game theory. While existing papers focus on the axiomatic motivation of Shapley values, and efficient techniques for computing them, they offer little justification for the game formulations used, and do not address the uncertainty implicit in their methods' outputs. For instance, the popular SHAP algorithm's formulation may give substantial attributions to features that play no role in the model. In this work, we illustrate how subtle differences in the underlying game formulations of existing methods can cause large differences in the attributions for a prediction. We then present a general game formulation that unifies existing methods, and enables straightforward confidence intervals on their attributions. Furthermore, it allows us to interpret the attributions as contrastive explanations of an input relative to a distribution of reference inputs. We tie this idea to classic research in cognitive psychology on contrastive explanations, and propose a conceptual framework for generating and interpreting explanations for ML models, called formulate, approximate, explain (FAE). We apply this framework to explain black-box models trained on two UCI datasets and a Lending Club dataset.
Luke Merrick, Ankur Taly
arXiv:1909.08128 · cs.LG, cs.AI, stat.ML · submitted Sep 17, 2019 · updated Jun 25, 2020
abstract · pdf · html
That reference paper is unifying in the sense that it could tie together a lot of frameworks such as LIME, QII, and others.
The paper in OP is unifying in the sense of the explanation games/game formulations (while also referencing QII).
The perspective of using contrastive explanations is very intuitive and becoming more common in different aspects of ML. My only wish is that I could try this FAE stuff myself. The authors of SHAP have a pretty immediately accessible python library. Could not find the one for this novel framework. It is not that I doubt the results. My point is a lot of good ideas like the ones presented in the paper often don't take off unless the authors also give easily pip-able libraries.
edit: I found the code[1], but it is not an easily conda-able library.
[1]: https://github.com/fiddler-labs/the-explanation-game-supplem...