Local Patterns: Theory and Practice of Constraint-Based Relational Subgroup Discovery

This paper investigates local patterns in the multi-relational constraint-based data mining framework. Given this framework, it contributes to the theory of local patterns by providing the definition of local patterns, and a set of objective and subjective measures for evaluating the quality of induced patterns. These notions are illustrated on a description task of subgroup discovery, taking a propositionalization approach to relational subgroup discovery (RSD), based on adapting rule learning and first-order feature construction, applicable in individualcentered domains. It focuses on the use of constraints in RSD, both in feature construction and rule learning. We apply the proposed RSD approach to the Mutagenesis benchmark known from relational learning and a real-life telecommunications dataset.