A Fast Algorithm for Frequent Itemset Mining Using Patricia* Structures

Abstract

Efficient mining of frequent itemsets from a database plays an essential role in many data mining tasks such as association rule mining. Many algorithms use a prefix-tree to represent a database and mine frequent itemsets by constructing recursively conditional prefix-trees from the prefix-tree. A (conditional) prefix-tree can be stored in various structures. The construction and traversal costs of prefix-trees, or rather their storage structures, take a large proportion in the whole cost for such algorithms. The PatriciaMine algorithm employs a Patricia trie to store a prefix-tree and shows good performance. In this study, we introduce an efficient Patricia* structure for storing a prefix-tree. A Patricia* structure is more compact and contiguous than a corresponding Patricia trie, and thus the construction and traversal costs of the former are less than those of the latter. Previous prefix-tree-based algorithms adopt a similar mining procedure, in which most nodes in a prefix-tree are repeatedly accessed when the prefix-tree is processed. The paper presents a novel mining procedure in which node accesses for a prefix-tree are greatly reduced. We propose the PatriciaMine* algorithm that is the combination of the Patricia* structure with the proposed procedure. Experimental data show that PatriciaMine* outperforms not only PatriciaMine but also several fast algorithms, such as FPgrowth* and dEclat, for various databases.