Providing User-Support in Performing Knowledge Discovery in Databases

Knowledge Management (KM) is becoming a success factor for industrial organisations. Obtaining control over and gaining information out of data helps to achieve the organisation’s goals more effectively. Thus knowledge (or information) becomes a very important resource. This resource must be adequately procured, stored, processed and communicated. These tasks are central points of Knowledge (and Information) Management which embraces key issues such as knowledge acquisition, data warehouses, data mining, data base management systems, knowledge representation, case-based reasoning, hyper media, workflow management, and decision support systems. As one can see, besides the information systems aspect AI methods and concepts play a significant role in Knowledge Management, esp. concerning procurement (knowledge acquisition), storing (knowledge representation), processing (e.g. data mining, decision support systems) knowledge. One important source of information and knowledge is the growing number of large databases which are built up and maintained in large organizations, but in the meantime also in a lot of small and medium size enterprises (SMEs). As consequence, the research and application area ’Knowledge Discovery in Databases’ (cf. [Frawley et al. 91], [Fayyad and Uthurusamy 95], [Simoudis et al. 96], [Fayyad et al. 96]) gained some importance during the last years. Successful KDD applications in marketing, financial investment, or network management indicate that the development of KDD applications may result in strategic advantages in performing the business tasks (cf. e.g. [Brachman et al. 96]). However, all experiences show that the development of (successful) KDD applications is a complex and errorprone process [Brachman and Anand 96]. In general, the KDD process consists of a task analysis step (for identifying the real application problem), a pre-processing step (for identifying and selecting relevant data), a data mining step, and a post-processing step (for evaluating the data mining results). At the moment, only first steps towards a comprehensive methodology that supports such complex and iterative KDD processes have been proposed (cf. e.g. [Wirth and Reinartz 96], [Engels 96]). On the other hand, there is a clear indication that such a methodology is needed since: ̄ especially in SMEs KDD applications will be developed by application specialists and not by KDD specialists, ̄ task analysis has to be supported in a systematic way in order to come up with a KDD problem specification which really meets the application needs, ̄ there exist so many dependencies between different KDD algorithms (e.g. for pre-processing and data mining) well as between data characteristics and applicable algorithms that a kind of planning support is required for achieving a well-defined and consistent KDD process, ̄ the KDD process is highly iterative which means that results of later steps (like e.g. evaluation) will have to fed back to earlier steps (like e.g. pre-processing). Our approach for providing user-support in performing Knowledge Discovery in Databases [Engels 96] aims at developing a methodology to support the user when performing entire KDD processes and to support the reuse of previously successfully applied KDD processes. The methodology should provide a user guidance module to enable application specialists (not necessarily KDD experts) to develop a KDD process by (re-)using successfully applied (parts of) KDD processes, data mining algorithms, and preand post-processing algorithms adequately. By supporting such a reuse-oriented approach the development time of new applications decreases while higher quality solutions are achieved. The user guidance module includes a repository which contains previously applied KDD processes, data mining algorithms, and pre/post-processing algorithms. To be able to retrieve these processes and algorithms (e.g. by applying case-based reasoning methods) there must exist a uniform description that highlights the differences in the functionality of these processes and algorithms. On the other hand, the initial