Research On Rough Set Based Methods For Data Mining And Decision Support

Recently, data mining has been one of hot research area. Data mining aims to extract implicit, previously unknown and potentially useful knowledge from large amount of data. The knowledge discovered by data mining technologies can be used to offer decision support. Rough set theory is an important mathematical tool for use in circumstances which are characterized by vagueness and uncertainty. It has been proven to be very useful in the field of data mining.The research work in this dissertation is based on the following observations: (1) most existing rough set methods lack of suitable means to deal with distributed data environment; (2) since the decision support ability of decision table will be reduced in rule induction process, the obtained rules can only offer limited decision support compared with the decision table. In order to resolve these two problems, meta-information method and rough decision support method are proposed respectively.Meta-information method aims to provide a cost-efficient means to deal with distributed data environment. Meta-information method consists of meta-information generation and maintenance method and rough set method based on meta-information. In meta-information method, the concept of information system hi original rough set theory is generalized by considering a family of subsystems. This kind of information system is called distributed information system. This generalization is natural to describe the distributed data sets in distributed data environment. A new concept, called meta-information, is proposed to describe the distributed information system or its subsystems. Meta-information eliminates the redundant data in the level of equivalent class, and describe the intersection between every pair of condition class and decision class in the form of number. Thus, compared with information system(or subsystems), meta-information requires far less storage, and its data structure is more simple and compact. Due to these properties, meta-information integration generally cost less than data integration, and the performance of methods based on meta-information is generally better than methods directly operating on original data.The meta-information generation method can generate meta-information incrementally for each subsystem, and produce the meta-information of distributed information system by integrating the meta-information of its subsystems. The meta-information maintenance method deals with the changed data in the underlying subsystems , and produces the up-to-date meta-information. Since meta-information is maintained in a dynamic way, the obtained meta-information can be saved for future reuse. The cost of meta-information integration and maintenance is much less than that of data integration and maintenance.Two rough set methods for relative attribute reduction and rule extraction aretnproposed on the basis of meta-information. These meta-information based methods can produce relative attribute reduct, and extract decision rules from information system, respectively. Due to the simpleness and compactness of meta-information, the time complexities of meta-information based rough set methods is generally lower than their non-meta-information versions. Besides, since the meta-information is reusable and can be maintained in a dynamic way, the rough set method can share the same procedure for data pre-processing(i.e., meta-information generation and maintenance), which will lessen the cost of data pre-processing.Rough decision support method is proposed to overcome the disadvantages described in the second observation. The rough decision support method is a family of rough set method for decision support, which include decision support methods based on condition vector, methods for additional condition vector acquisition, and methods for condition vector reduction. These methods can find a preferred decision related to some condition vector or similar condition vector, increase the condition vector's decision strength to some decision vector b...