Counterfactuals, Bayesian Network, Explainability, Probabilistic Graphical Model, Explainable AI
Abstract
With the advances in the Artificial intelligence field, methodologies to explain the decision from models and their reasoning started gaining attention in order to increase the system's trustworthiness for the user and the decision-maker. Among these methods, is the explainability by counterfactuals, which states a question of a hypothetical different world that the model would produce a different and desired output that is closest possible to the observed world, which is linked to the usual concept of causality in statistics. In this paper, we present a method to use counterfactuals to explain the Bayesian Network, a probabilistic graphical model method, results for given evidence, by finding the best possible closest imaginary worlds in which the result was equal to a desired outcome, different from the observed output. The proposed methodology guarantees the plausibility of the counterfactual scenarios, by only changing feasible pieces of evidence, minimal variable state changes, and providing metrics in terms of diversity of generated data by returning all the possible scenarios to receive the specific outcome ranked by the system's confidence. Moreover, this method presents the benefits of using probabilistic models, such as performing inference over incomplete datasets and seeing the output's confidence in the output for each hypothetical evidence set.
