GX-HUI: Global Explanations of AI Models based on High-Utility Itemsets

Shapley Values are established concepts used to explain local and global contribution of individual features to the prediction of AI models. Currently, global Shapley-based explainers do not consider the co-occurrences of feature-value pairs in the analyzed data. This paper proposes a novel approach to leverage the High-Utility Itemset Mining framework to jointly consider Shapley-based feature-level contributions and feature-value pair co-occurrences. The results achieved on benchmark datasets show that the extracted patterns provide actionable knowledge, complementary to those of global Shapley Values.