Signature-based Tree for Finding Frequent Itemsets

Abstract

The efficiency of a data mining process depends on the data structure used to find frequent itemsets. Two approaches are possible: use the original transaction dataset or transform it into another more compact structure. Many algorithms use trees as compact structure, like FP-Tree and the associated algorithm FP-Growth. Although this structure reduces the number of scans (only 2), its efficiency depends on two criteria: (i) the size of the support (small or large); (ii) the type of transaction dataset (sparse or dense). But these two criteria can generate very large trees. In this paper, we propose a new tree-based structure that emphasizes on transactions and not on itemsets. Hence, we avoid the problem of support values that have a negative impact on the generated tree.