Knowledge Discovery in Database Using Concept Learning with Background Knowledge

Abstract

This paper proposes a new approach to discover knowledge hidden in a database. The approach is based on theories of concept formation, which is a type of concept learning that learns classification of data without a priori knowledge about which data is classified to which class. However, conventional theories of concept formation proposed so far have some crucial problems to be applied to a large amount of data in a database. Therefore, we introduces a new type of concept formation that can be applied to databases. Our approach is featured in the following aspects of learning ; 1) non-incremental learning, 2) learning with background knowledge about attributes of the database, and 3) learning by data both with nominal and with numerical attributes. The conventional theories of concept formation, on the other hand, employ incremental learning without any background knowledge through data with either nominal or numerical attributes. Since they assume that data comes one by one, they must incorporate, create, merge, or split concepts whenever a data comes. It is inefficient, however, when all data are given at once. So, our approach generates concepts in top down direction in the hierarchy using the background knowledge about attributes of data in the database. This becomes possible, because it assumes that all data are given at once. Both the background knowledge and the top down strategy realize efficient concept formation from a database. Employment of background knowledge also gives the capability to express the generated concepts in labels used in the knowledge. Then, concept formations including our approach employ a criterion to measure the quality of each generated concept. Hierarchical concepts are generated using this criterion. However, conventional approaches use a criterion which is defined only for nominal attributes or for numerical ones. Such restriction is too unrealistic to apply to ordinary databases both with nominal and numerical attributes. Therefore, our approach introduces a new criterion that can deal with the both attributes. Finally, the proposed approach is applied to a real estate database, and its effectiveness is shown.