Pruning relations for substructure discovery of multi-relational databases

Multirelational data mining methods discover patterns across multiple interlinked tables (relations) in a relational database. In many large organizations, such a multi-relational database spans numerous departments and/or subdivisions, which are involved in different aspects of the enterprise such as customer profiling, fraud detection, inventory management, financial management, and so on. When considering multirelational classification, it follows that these subdivisions will express different interests in the data, leading to the need to explore various subsets of relevant relations with high utility with respect to the target class. The paper presents a novel approach for pruning the uninteresting relations of a relational database where relations come from such different parties and spans many classification tasks. We aim to create a pruned structure and thus minimize predictive performance loss on the final classification model. Our method identifies a set of strongly uncorrelated subgraphs to use for training and discards all others. The experiments performed demonstrate that our strategy is able to significantly reduce the size of the relational schema without sacrificing predictive accuracy.