A high-performance algorithm for frequent itemset mining

Abstract

Frequent itemsets, also called frequent patterns, are important information about databases, and mining efficiently frequent itemsets is a core problem in data mining area. Pattern growth approaches, such as the classic FP-Growth algorithm and the efficient FPgrowth*algorithm, can solve the problem. The approaches mine frequent itemsets by constructing recursively conditional databases that are usually represented by prefix-trees. The three major costs of such approaches are prefix-tree traversal, support counting, and prefix-tree construction. This paper presents a novel pattern growth algorithm called BFP-growth in which the three costs are greatly reduced. We compare the costs among BFP-growth, FP-Growth, and FPgrowth*, and illuminate that the costs of BFP-growth are the least. Experimental data show that BFP-growth outperforms not only FP-Growth and FPgrowth*but also several famous algorithms including dEclat and LCM, ones of the fastest algorithms, for various databases. © 2012 Springer-Verlag.