A parallel SAT-based framework for closed frequent itemsets mining

Constraint programming (CP) and propositional satisfiability (SAT) based framework for modeling and solving pattern mining tasks has gained a considerable audience in recent years. However, this nice declarative and generic framework encounters a scaling problem. The huge size of constraints networks/propositional formulas encoding large datasets is identified as the main bottleneck of most existing approaches. In this paper, we propose a parallel SAT based framework for itemset mining problem to push forward the solving efficiency. The proposed approach is based on a divide-and-conquer paradigm, where the transaction database is partitioned using item-based guiding paths. Such decomposition allows us to derive smaller and independent Boolean formulas that can be solved in parallel. The performance and scalability of the proposed algorithm are evaluated through extensive experiments on several datasets. We demonstrate that our partition-based parallel SAT approach outperforms other CP approaches even in the sequential case, while significantly reducing the performances gap with specialized approaches.