Exploiting the Duality of Maximal Frequent Itemsets and Minimal Infrequent Itemsets for I/O Efficient Association Rule Mining

Abstract

Any algorithm for mining association rules must discover the set of all maximal frequent itemsets (maxL) from a database. Given a set of itemsets X, to verify that X is maxL, two conditions must be checked: (1) any itemset x in X is frequent, and (2) the dual of X must be the set of all minimal infrequent itemsets (minS). This observation leads us to a family of algorithms for mining association rules. Given a reasonable guess of minS and maxL, we verify their duality relationship, and refine the two sets until the above two conditions hold. We note that previously proposed algorithms such as Apriori and Pincer-Search are all members of our algorithm family. Also, we study a member algorithm called FlipFlop. Through a series of experiments, we show that FlipFlop significantly reduces the I/O requirement of mining association rules.