Towards explainable artificial intelligence with potential games
Evangelos D. Spyrou, Vassilios A. Kappatos, Afroditi Anagnostopoulou, Evangelos D. Bekiaris · Mathematical Models in Engineering · 2025
Explainable Artificial Intelligence (XAI) emerged when researchers attempted to identify methods that would interpret the models that are used to perform classification and predictions, in order to avoid having a black box just informing about the result. Methods of XAI are crucial to determine details of the model feature contribution towards the result. One of these methods is attributed to cooperative game theory and especially Shapley values. With this method the features are considered as players and the marginal contribution of the features are employed. In this paper, we take onboard the Potential Game paradigm to show the interconnection between them and the Shapley values. We show that the Shapley values are interlinked with the potential function. Moreover, we setup a game with the marginal contribution of the players as their utility functions and we prove that the game is a potential game. Finally, we show that the price of stability of this game is 1. We utilise the Simulated Annealing (SA) method to find the optimal solution.