On discovering "potentially useful" patterns from databases

Abstract

As is generally accepted, the most important feature that a KDD system must possess is the ability to discover patterns that are "potentially useful". In order to allow KDD systems to make potentially useful judgments, we give formal definitions of "potential usefulness" by completely staying within the realms of the expressiveness provided by Bacchus' Probabilistic Logic Language. Furthermore, a tractable algorithm is proposed that is capable of discovering all potentially useful patterns from databases, given limited accessible information.