Research On Data Stream Classification Based On Granular Computing And F-Rough Sets Extension

Data stream is the main form of big data. Concept drifting detection and data stream classification are dominant in data stream mining. There are many methods to concept drifting detection, but they have some drawbacks, such as reducing redundant attributes lonely in every sliding window, and detecting concept drifting with outer properties, etc.Rough sets can effectively deal with imprecise, incomplete and imperfect information and knowledge, which do not need any priory knowledge, and directly discover implicit knowledge and potential rules through data analysis and reasoning. It is difficult for classical rough set theory which can’t suitable for the massive research, dynamic data, or data stream. F-rough sets extends Pawlak rough set theory from one single decision table to a family of decision tables, which can investigate and analyze the change of data from the wholeness and parts. It is a potential power to be a strong theoretical and practical tool for analyzing data stream. However, F-rough sets are limited, when continuous data are handled.Our work in this thesis mainly contains two parts:(1)Based on the basic principles of rough sets and F-rough sets, sliding windows in a data stream are regarded as decision subsystems, and the attribute significance of conditional attributes is employed to detect concept drifting. (2)Combined with F-rough sets and fuzzy rough sets, F-fuzzy rough set model is introduced, and the method of its attribute reduction is proposed.The main points of innovation are showed as follows:(1) Concept drifting is defined with attribute significance.(2) Redundant attribute are deleted with parallel reducts, which unifies the measurement of concept drifting.(3) Concept drifting is detected by internal properties.(4) The model of F-fuzzy rough sets is introduced, and its reduction is presented.