AqBC: a multistrategy approach for constructive induction

In order to obtain potentially interesting patterns and relations from large, distributed, heterogeneous databases, it is essential to employ an intelligent and automated KDD (Knowledge Discovery in Databases) process. One of the most important methodologies is an integration of diverse learning strategies that cooperatively performs a variety of techniques and achieves high quality knowledge. AqBC is a multistrategy knowledge discovery approach that combines supervised inductive learning and unsupervised Bayesian classification. This study investigates creating a more suitable knowledge representation space with the aid of unsupervised Bayesian classification system, AutoClass. AutoClass discovers interesting patterns from databases. Via constructive induction, these patterns modify the knowledge representation space so that the robust inductive learning system, AQ15c, learns useful concept descriptions of a taxonomy. AqBC applied to two different sample problems yields not only simple but also meaningful knowledge due to the systems that implement its parent approaches. AqBC´s good performance appears to be due to its integration of reliable unsupervised Bayesian classification, constructive induction and rule induction, and not to the presence of any component alone