Efficiently Mining Maximal Diverse Frequent Itemsets

Given a database of transactions, where each transaction is a set of items, maximal frequent itemset mining aims to find all itemsets that are frequent, meaning that they consist of items that co-occur in transactions more often than a given threshold, and that are maximal, meaning that they are not contained in other frequent itemsets. Such itemsets are the most interesting ones in a meaningful sense. We study the problem of efficiently finding such itemsets with the added constraint that only the top-k most diverse ones should be returned. An itemset is diverse if its items belong to many different categories according to a given hierarchy of item categories. We propose a solution that relies on a purposefully designed index structure called the FP*-tree and an accompanying bound-based algorithm. An extensive experimental study offers insight into the performance of the solution, indicating that it is capable of outperforming an existing method by orders of magnitude and of scaling to large databases of transactions.