Research On Approaches Of Dynamic Attribute Reduction Based On Knowledge Granularity

With the fast development of computer network technology, computer storage technology and communication technology, the rapid growth big data provides the opportunities and challenges for enterprises. How to acquiring knowledge effectively and timely from the big data has become an urgent problem. In addition, the decision system varies dynamically with time, e.g., the old data is deleted, new data is added and the error data is revised in real-life applications. How to update knowledge effectively has become a key research topic in the area of information sciences. Granular computing provides a multi-level, divide and conquer solution to process dynamic data and knowledge discovery efficiently. Based on the theory of granular computing, rough set and the incremental learning technology, this dissertation focuses on the development of dynamic attribute reduction methods when the decision system varies dynamically. The main research achievements are summarized as follows.(1) How to effectively carry out attribute reduction is an important task when some objects vary dynamically in the decision system. The matrix-based incremental mechanisms for knowledge granularity are discussed and the corresponding matrix-based dynamic attribute reduction algorithm is proposed when some objects vary dynamically. However,matrix-based dynamic attribute reduction methods are only suitable for the small data sets.Incremental mechanisms based on non-matrix for knowledge granularity are further analyzed and the dynamic attribute reduction algorithm based on non-matrix for updating reduct is developed when some objects vary dynamically.(2) How to effectively carry out attribute reduction is an important task when multiple attributes are added into the decision system. The matrix-based incremental mechanisms for knowledge granularity are analyzed and the corresponding matrix-based dynamic attribute reduction method for updating reduct is developed when multiple attributes are added into the decision system. However, matrix-based dynamic attribute reduction methods are only suitable for the small data sets. Incremental mechanisms based on non-matrix for knowledge granularity are further discussed and the dynamic attribute reduction algorithm based on non-matrix is proposed when multiple attributes are added into the decision system.(3) How to effectively carry out attribute reduction is an important task when attribute values of an object have been replaced by new ones in the decision system. The incremental mechanisms for knowledge granularity are analyzed and the corresponding dynamic attribute reduction approach for updating reduct is proposed. When attribute values of multiple objects have been replaced by new ones in the decision system, the algorithm is very time-consuming to find reduct since it has to be executed repeatedly. Incremental mechanisms with varying attribute values of multiple objects for knowledge granularity are further discussed and the corresponding dynamic attribute reduction algorithm is developed.(4) How to effectively carry out attribute reduction is an important task when data sets are in a large scale. The mechanisms for reduct of large-scale data set are analyzed with a multi-granulation view and divide and conquer solution. An attribute reduction algorithm is developed with a multi-granulation view. Then, the incremental mechanisms for knowledge granularity are analyzed and the corresponding dynamic attribute reduction approaches for updating reduct are developed with a multi-granulation view when some objects vary dynamically.In this dissertation, based on the theory of rough sets, granular computing and the incremental mining techniques, the incremental approaches for attribute reduction based on knowledge granularity are developed. Several incremental mechanisms are presented for data mining and knowledge discovery when the decision system varies dynamically.Experiments conducted on different data sets from UCI have verified the efficiency and effectiveness of the proposed incremental algorithms. The achievements of this study can not only be applied to dynamic attribute reduction from massive data, but also provide analysis of big data with theories, methodologies and algorithms together with the extension of the application fields of rough sets and granular computing theory.