Why Shapley Value and Its Variants Are Useful in Machine Learning (and in Other Applications)
摘要
Shapley value—a useful way to allocate gains in cooperative games—has been very successful in machine learning (and in other applications beyond cooperative games). This success is somewhat puzzling, since the usual derivation of the Shapley value is based on requirements like additivity that are natural in cooperative games and but not in machine learning. In this paper, we provide a new simple derivation of the Shapley value, a derivation that does not use game-specific requirements like additivity and is, thus, applicable in the machine learning case as well.