Knowledge-Intensive, Interactive and Efficient Relational Pattern Learning

Abstract

Abstract : This report describes the work to develop and evaluate state-of-the-art relational pattern learning algorithms for the Evidence Assessment, Grouping, Linking and Evaluation (EAGLE) program. The University of Wisconsin team consisted of leaders in relational data mining and relational machine learning, in particular inductive logic programming (ILP). Major contributions by the team included the development of an ILP system implemented entirely in database operations, FOIL-D, and a statistical relational learning (SRL) system that incorporates explicit probabilistic constraints into ILP, CLP(BN). CLP(BN) incorporates all the representational power of probabilistic relational models (PRMs) but uses an ILP approach to learning. Another major contribution is the definition and development of View Learning, an approach to change of representation for SRL. Though SRL systems are particularly well-suited to the goals of EAGLE - indeed, the field of SRL received much of its impetus for growth from EAGLE - these systems have been constrained to work with the input representation, typically a relational schema. View Learning in SRL permits the definition of a new schema - a new view of the database - better suited to the learning goals. This project also made advances within learning ensembles, including the DECORATE approach to diverse ensembles, the use of bagging within ILP, the GLEANER algorithm to construct ensembles of relational rules having varying trade-offs of precision and recall, and a parallel implementation of bagging in ILP. The project also contributed novel stochastic search algorithms for ILP.