An improved algorithm for mining frequent Inter-transaction patterns

Abstract

Mining Inter-transaction patterns (ITPs) from large databases is a common data mining task, which discovers the patterns across several transactions in a transaction database. Although, several algorithms have been proposed for this task, they remain computationally expensive. To resolve this issue, this paper presents an efficient method called DITP-Miner to mine ITPs. In our proposed algorithm, there are four phases. First, we scan the database once to find frequent 1-patterns with their tidsets. Second, we generate inter-transaction 1-pattern candidates with the given span values and sort all the frequent 1-patterns in an ascending order according to their supports. Third, based on frequent items found in phases 1 and 2, we find frequent 2-patterns with their diffsets. In the fourth phase, we use diffsets and DFS (Depth-First-Search) technique to get all frequent ITPs. In addition, three propositions are also offered to early prune infrequent patterns in the processing. Proposition 1 is used to early prune infrequent inter-transaction 1-pattrens, Proposition 2 is used to quickly compute the support of patterns, and Proposition 3 (subsume concept) is used to quickly compute the support of patterns and to early prune infrequent patterns, which reduce the search space. Through experimental results, we find out our proposed approach is more efficient than ITP-miner in both the mining time and the memory usage.