Aggregation-Based Feature Invention and Relational

Due to interest in social and economic networks, relational modeling isattracting increasing attention. The field of relational datamining/learning, which traditionally was dominated by logic-basedapproaches, has recently been extended by adapting learning methods suchas naive Bayes, Baysian networks and decision trees to relational tasks.One aspect inherent to all methods of model induction from relationaldata is the construction of features through the aggregation of sets.The theoretical part of this work (1) presents an ontology of relationalconcepts of increasing complexity, (2) derives classes of aggregationoperators that are needed to learn these concepts, and (3) classifiesrelational domains based on relational schema characteristics such ascardinality. We then present a new class of aggregation functions, onesthat are particularly well suited for relational classification andclass probability estimation. The empirical part of this paperdemonstrates on real domain the effects on the system performance ofdifferent aggregation methods on different relational concepts. Theresults suggest that more complex aggregation methods can significantlyincrease generalization performance and that, in particular,task-specific aggregation can simplify relational prediction tasks intowell-understood propositional learning problems