Playing the network backward: A Game Theoretic Attribution Framework
This research recasts model attribution methods, like gradients and LRP, as a two-player game on a network graph. It introduces games—a Stopping Game for gradients and an Entropy-Regularised Routing Game for LRP—to provide a unified framework. This allows researchers to compare different backward calculations rigorously using metrics like Hellinger distance, enabling the specification of desired explanation properties through game theory.