The Research Of Parallel Reduction Based On Rough Set

Rough set theory, proposed by Poland scholar Pawlak in 1982, is a powerful tool for data mining and an important branch of Granular computing. The rough set theory provides a systematic method for dealing with imperfect and insufficiency information by indistinguishable relationship. Rough set has become a hot topic in computer science, information science and artificial intelligence etc.Knowledge reduction is an important application and a core issue in rough set theory. Knowledge reduction deletes the redundancy information in information systems or decision tables while it does not change the classification of whole dataset. Because of the complexity of the dataset and the localization of traditional reductions, rough set can't show its vitality while we deal with the mass data. Many scholars make an attempt on improvement of knowledge reduction algorithm, such as parallel reduction.Parallel reduction has become a hot topic in recent years. It makes use of the idea and method of parallel computing, and can be used in data mining that based on rough set. Because Parallel reduction is a newly concept, its paradigm and method and technology should be improved. But now there isn't an efficient Parallel reduction method. The research of this thesis is mainly about the parallel reduction definition, properties based on rough set and sub-table sampling for different tables.The main contributions and innovations of this thesis are as follows. At first, according to the characteristic of different data, we introduced the definition of parallel reduction, and extend its concept. We introduced three types parallel reduction:parallel reduction based on positive area, parallel reduction based on discernibility matrix, parallel reduction based on importance degree of attribute. Meanwhile we designed three algorithms for the three parallel reductions.Secondly, because of the stringent of the parallel reduction based on positive area, we proposed the definition of parallel reduction based on variable precision, and investigated its properties.Thirdly, we proposed different sampling strategies based on different date set; we also proposed a clustering method for sub-table extraction.Finally, we analyzed the limits of decision rules extraction, and proposed an effective incremental extraction of decision rules to solve these problems.