| Data mining is very useful for gleaning insight about entities that can be described by, have, generate, or otherwise deal with large volumes of data. Traditional methods discover and present acquired knowledge not only as relationships between two attributes, but often between sets of attributes. However, further knowledge can be acquired by considering a discovered fact as something more than a set or a sequence. As a set, a fact is simply a flat list of items, where ordering is sometimes not important. Allowing a fact to be a structured entity, such as a graph, makes it a more informative medium than if it just were a set. Many natural and artificial objects are inherently structured and complex. Data is acquired from operation of a factory, including sensors and other mechanical objects. Database operations (transactions, query subsystems), computer systems (file systems, network operations, program traces, performance data), medical issues (temporal and other relationships among symptoms, illnesses and treatments), and other sources are also rich in structure. This data can be processed using the proposed methodology to reveal interesting and unexpected relationships and behavior, which would go unnoticed, were it only processed using queries or other data mining methods that are not designed to handle structured data. The methodology for the automated discovery of such structured facts in arbitrary multigraphs (not just transactionally partitioned) using a guided search of the entire associated combinatorial space is proposed and studied. Benefits, limitations, and other associated issues are discussed. |
