STATISTICAL, LOGIC BASED, AND NEURAL NETWORK BASED METHODS FOR MINING RULES FROM DATA

摘要

This lecture has given an overall survey of the area of data mining, emphasizing the need for integrative frameworks that allow to apply different data mining methods in a common context and to view them from a common perspective. An integrative framework based on observational logic is described, and its use is explained for two extremely important kinds of data mining methods--statistical methods based on hypotheses testing, and methods for the extraction of rules from data with artificial neural networks. The calculus of obsrvational logic has been originally proposed as a means for a unified treatment of different kinds of statisical methods. In this role, it has been more than 25 years used in practice, in the method GUHA, one of the oldest and most vital methods for automated knowledge discovery. The lecture tried to show that the presented framework provides a unique conceptual view of both kinds of considered data mining methods while preserving their specific advantages. Due to the different underlying paradigms and different initial assumptions, artificial neural networks do not necessarily yield the same results as statistical methods. Hence, a coincidence between relationships discovered in the data by both kinds of methods increases the chance that those relationships pertain also to the reality behind the data, to the phenomena that generated them. On the other hand, if a relationship found by means of some quantifier can not be confirmed by means of other quantifiers, including a quantifier based on the other paradigm, then such as relationship is a natural starting point for further, deeper investigations.