On Maximal Frequent Itemsets Enumeration

Abstract

Enumerating interesting patterns from data is an important data mining task. Among the set of possible relevant patterns, maximal frequent patterns is a well known condensed representation that limits at least to some extent the size of the output. Recently, a new declarative mining framework based on constraint programming (CP and satisfiability (SAT) has been designed to deal with several pattern mining tasks. For instance, the itemset mining problem has been modeled as a constraint network/propositional formula whose models correspond to the pattern to be mined. In this framework, closeness, maximality and frequency properties can be handled by additional constraints/formulas. In this paper, we propose a new propositional satisfiability based approach for mining maximal frequent itemsets that extends the one proposed in [13]. We show that instead of adding constraints to the initial SAT based itemset mining encoding, the maximal itemsets, can be obtained by performing clause learning during search. Our approach leads to a more compact encoding. Experimental results on several datasets, show the feasibility of our approach.