Summary: Generalized itemset mining is a powerful tool to discover multiple-level correlations among the analyzed data. A taxonomy is used to aggregate data items into higher-level concepts and to discover frequent recurrences among data items at different granularity levels. However, since traditional high-level itemsets may also represent the knowledge covered by their lower-level frequent descendant itemsets, the expressiveness of high-level itemsets can be rather limited. To overcome this issue, this article proposes two novel itemset types, called Expressive Generalized Itemset (EGI) and Maximal Expressive Generalized Itemset (Max-EGI), in which the frequency of occurrence of a high-level itemset is evaluated only on the portion of data not yet covered by any of its frequent descendants. Specifically, EGI s represent, at a high level of abstraction, the knowledge associated with sets of infrequent itemsets, while Max-EGIs compactly represent all the infrequent descendants of a generalized itemset. Furthermore, we also propose an algorithm to discover Max-EGIs at the top of the traditionally mined itemsets.Experiments, performed on both real and synthetic datasets, demonstrate the effectiveness, efficiency, and scalability of the proposed approach.