Strategies for partitioning data in association rule mining

The problem of extracting association rules from databases is well known. The most demanding part of the problem is the determination of the support for all those sets of attributes which occur often enough to be of possible interest. We have previously described methods we have developed that approach the problem by first constructing a tree (the P-tree) that contains a record of all the relevant information in the database and a partial computation of the support totals. This approach offers significant performance advantages over comparable alternative methods, which we have demonstrated experimentally with store-resident datasets. In practice, however, the real focus of interest is on much larger databases. In this paper we discuss strategies for partitioning the data in these cases, and present results of the performance analysis.