Mining Frequent Patterns Based on Data Characteristics

Frequent pattern mining is crucial part of association rule mining and other data mining tasks with many practical applications. Current popular algorithms for frequent pattern mining perform differently: some are good for dense databases while the others are ideal for sparse ones. In our previous research, we developed a new frequent pattern mining algorithm named FEM that runs fast on both sparse and dense databases. FEM combines the mining strategies of FP-growth and Eclat and given a userspecified threshold it adapts its mining behaviors to the data characteristics to efficiently find all short and long patterns from different database types. However, for best performance of FEM, an appropriate threshold value used to control the switching between its two mining tasks need to be selected by the user. In this paper, we present DFEM, an improved algorithm of FEM that automatically adopts a runtime dynamic threshold to better fit to the characteristics of the databases. The experimental results show that DFEM outperforms FEM and other popular frequent pattern mining algorithms including Apriori, Eclat, FP-growth on both sparse and dense databases.