A hybrid approach to rule discovery in databases

Abstract

This paper introduces a hybrid approach for rule discovery in databases in an environment with uncertainty and incompleteness. We first create an appropriate relationship between deductive reasoning and stochastic process, and extend the relationship for including abduction. Then, we define a Generalization Distribution Table (GDT), which is a variant of transition matrix in stochastic process, as a hypothesis search space for generalization, and describe that the GDT can be represented by knowledge-oriented networks. Furthermore, we describe a discovery process based on the network representation. Finally, we introduce some extension for making our approach more useful, and discuss some problems for real applications. We discuss inductive methods from the viewpoint of the value of information, and describe that the main features of our approach are: (1) the uncertainty of a rule, including its ability to predict possible instances, can be explicitly represented in the strength of the rule, (2) noisy data and data change can be handled effectively, (3) biases can be flexibly selected and background knowledge can be used in the discovery process for constraint and search control, and (4) if-then rules can be discovered in an evolutionary, parallel-distributed cooperative mode.