Generating Patterns for Knowledge Discovery using First Principles Modeling of Activity

Knowledge discovery algorithms generate alerts to interesting activity by finding partial matches to user-specified patterns of interest. Several statistical models can be used to generate pattern based on training data. However, when training sets are not readily available it is still possible to construct patterns from domain knowledge. To test the feasibility of such an approach we examine a database of publications in biological sciences, BioBase, and attempt to predict whether or not two researchers who did not coauthor a paper in 1998-2002 will coauthor a paper in 2003. We constructed a set of four distinct patterns depicting scenarios in which a new coauthoring relationship might emerge. In each of the scenarios, we needed to identify individuals playing different roles within research groups, and use the number of publications of individuals to predict those roles. We also use the 1998-2002 coauthor relationships to provide evidence of the associated research groups and collaborations. Our results in a test database show that this approach is feasible and competitive when compared to others that rely on more extensive statistical modeling