Data mining algorithms look for patterns in data, traditionally represented in a single set. However, when mining for properties of compound objects (for example, logs and financial or scientific data), more adequate representation of these structures is needed. As the source of the data is often a relational database, it is important to extend previous mining methods to deal with objects stored in multiple related tables. This is the target of the multirelational data mining discipline.

Several approaches have been developed in recent years to achieve the tight coupling of data mining techniques to relational database management system (RDBMS) technology. These approaches include solutions provided by both commercial companies and academic research groups. The first approaches simply mapped multiple tables into one before analysis; the results were redundant and sometimes hid the structures of the compound objects.

The authors of this paper extracted the tree-like structures of objects stored in a database in different related tables, allowing existing tree-mining techniques to be applied in RDBMS. (The issue of how to deal with tables embodying M:N relationships was left open in the paper.) They show two representations, and the essence of the trees is extracted in two ways. The resulting 2x2 variants are compared both informally and formally.

The authors show how to determine frequent patterns within the set of trees, and how to mine for association rules among them, filtered by three kinds of rule-mining constraints. Finally, experimental results are shown for different datasets, using the authors’ tree-mining system.

I recommend this paper to interested data mining researchers.