Using Data Cubes for Metarule-Guided Mining of Multi-Dimensional Association Rules

Metarule-guided mining is an interactive approach to data mining, where users probe the data under analysis by specifying hypotheses in the form of metarules, or pattern templates. Previous methods for metarule-guided mining of association rules have primarily used a transac-tion/relation table-based structure. Such approaches require costly, multiple scans of the data in order to nd all the large itemsets. In this paper, we employ a novel approach to metarule-guided, multi-dimensional association rule mining which explores a data cube structure. We propose algorithms for metarule-guided mining using diierent cube structures: given a metarule containing p predicates, we compare mining on an n-dimensional cube structure (where p < n) with mining on smaller multiple p-dimensional cubes. In addition, we propose an eecient method for precomputing the cube, which takes into account the constraints imposed by the given metarule. Our preliminary performance study shows that a cube-based metarule-guided algorithm for mining multi-dimensional association rules is eecient and can easily be extended to the mining of multi-level, multi-dimensional association rules.