| Several problems exist for data mining in real world databases: the difficulty of determining a decision attribute when limited domain knowledge exists, the difficulty in selecting a decision attribute from a new table formed from joining several relations, and the problem of elaborate data selection in the knowledge discovery process. This thesis presents research results to solve these problems using methods that (1) determine a good decision attribute based on an approach developed from rough set theory and decision tree generation, (2) find meaningful frequent patterns based on attributes and dependencies in relational databases which are used to generate tables, and (3) discover optimal descriptive rule sets with respect to a users' interests, which copes with high dimensional and voluminous data sets, and which may contain numerous manifest facts. A stepwise refinement technique relying on the users' domain knowledge and interests has been devised for interactive refinement. This technique takes advantage of the fact that more general concepts occur more frequently and the focus of knowledge discovery is to find some hidden information that govern substantial portion of the database. |
