Research On Multivariate Information Granulation And Attribute Selection Approaches

With the rapid development of data acquisition,transmission and storage technology,the corresponding description,characterization and record of one object becomes more abundant,comprehensive and enduring,respectively.Accordingly,data expresses a typical characteristic of high-dimensional.As an effective method in data mining and knowledge discovery,attribute reduction has been wildly promoted and employed to deal with high-dimensional data.Nevertheless,in today's era of Big Data,the complex characteristics in data are not only limited in high-dimensional,but also reflected in weak supervision,multi-scale,etc.Therefore,faced with these complex characteristics,how to conduct information processing and analysis efficiently and accurately has become the key problem in attribute selection approach.In view of the complex characteristics in data in real-world application,in this dissertation,the information granulation and hierarchical cognitive model of human thinking and solving complex problems is adopted,the intended target is analyzed with multi-perspective,it follows that supervised information granulation and fusion,semi-supervised rough data analysis,multi-granularity attribute selection are discussed and studied thoroughly,mainly including: develop the multi-granulation and fusion for complex data,ensemble uncertainty data mining and analysis technology,and construct a set of attribute selection frameworks from the viewpoint of multi-granularity.Consequently,these researches can improve the performance of learners driven by complex data and the time efficiency of problem solving effectively.Specifically,the main contributions of this dissertation mainly and our innovations include the following aspects.1.A double-radius based multiple supervised neighborhood information granulation is proposed.With the reviewing of researches of Granular Computing,to realize the information granulation,most of the methods are actually contained in unsupervised learning,and the strong supervision is usually ignored.However,these methods may lack generalization performance in supervised learning tasks.To fill such a gap,by introducing the pairwise constraint,samples are divided between classes in view of the label information,and the mechanism of intra and extra double radii is designed to regulate the similarities between the samples in different groups,respectively.Correspondingly,the supervised neighborhood information granulation and reconstruction has been given.Such approach may filter the imprecise and inconsistent information remained in the process of information granulation,and it may effectively improve the anti-interference ability of information granulation with attribute noises.2.A semi-supervised ensemble based rough uncertainty analysis method is proposed.With the reviewing of rough uncertainty analysis,most of the methods actually focus on the decision systems with complete labels,and rely on the interactive results between attributes and labels.Additionally,such method is not enough to explore underlying knowledge and rules.Therefore,it may be useless when handling data with missing labels.To fill such a gap,the weakness of the existing rough uncertainty analysis methods in the semi-supervised problem is discussed in detail,and a novel strategy of ensemble based semi-supervised rough uncertainty analysis is proposed.Based on the local perspective,such approach can not only offer the semi-supervised attribute selection with a reasonable semantic explanation,but also widens the application prospect of rough set method in this problem.3.A framework of multi-granularity attribute selection is proposed.With the reviewing of attribute reduction,most of the methods actually take one and only granularity into account,and they may be useless in the granularity diversity caused by data disturbance.To fill such a gap,the construction process of attribute reduction based on the concept of granularity is revealed,and the internal correlation mechanism between information granulation and attribute reduction is clarified.Furthermore,the definition of multi granularity attribute reduction is given from the viewpoint of information fusion.Finally,and a general framework of efficient multi-granularity attribute selection algorithms is designed.Such approach can effectively reduce the limitations of attribute reduction in multi-granularity problem,and significantly improve the time efficiency of deriving reducts.