| The research of data mining and machine learning mainly focuses on how to acquire valid information and knowledge from the complex big data.However,statistical machine learning-based intelligent technologies merely emphasize the mining and understanding of low-level features of data from the perspective of data itself,which neglects the common features of uncertainty in the physical world and the cognitive process,resulting in the problems of separate expression between data and knowledge and contradiction between information process mechanism and human cognitive mechanism.Multi-granular cognitive computing is a kind of computing model that simulates human hierarchical cognitive mechanism and realizes the multi-granularity joint problem-solving mechanism of complex problems by constructing the association between different granularity layers or different nodes in the same granularity layer.Compared with statistical machine learning,multi-granular cognitive computing models are preferred to express a wide variety of forms of uncertainty phenomenon,such as the ambiguity of characterizing the concept,the impreciseness of data,and the inconsistency of knowledge representations among various cognitive levels,thus,is more suitable for processing the tasks of expression,measurement,and solution of knowledge in complex big data under the real world.Currently,the multi-granular cognitive computing models represented by quotient space,fuzzy set,and rough set have been deeply studied uncertainty problems in various scenarios,and have achieved many fruitful works.However,these works are always limited to a single form of uncertainty,a single granular layer,or a single granular space,and neglect the feature of the coexistence of diverse uncertainties involved in a multi-granular knowledge space.Hence,the current uncertainty measures are unable to accurately reflect similarity between different concepts and the difference of a same concept at various granular layers,thus,are unable to accurately represent multi-granular features of the target concepts with a complex structure such as hierarchical semantic labels.To address the above issues,this thesis analyzes the relationship between different forms of uncertainties in the bidirectional cognitive process between data and knowledge,and proposes the similarity measure based on the combination of concept extension and intension.Then,this thesis analyzes the difference of a concept at various granular layers in the multi-granular knowledge space,and proposes the corresponding uncertainty measures.On this basis,through granulating hierarchical classification target concept by knowledge distance of quotient space,this thesis constructs the multi-granular knowledge space satisfying the combination of concept extension and intension,and proposes the approximate description of the double granular space structure induced by target concept and condition attribute granular space.For discrete and continuous-valued condition attribute decision systems,this thesis respectively discusses the mechanism of granularity layer changes in double granular space from the perspectives of information entropy and cost measure,and proposes the corresponding granularity optimization methods.To sum up,the following aspects of this thesis are carried out:(1)Based on randomness and fuzziness compounded uncertainty form of cloud model,the concept of uncertainty distribution of cloud drops is proposed,which can realize the similarity measure based on the combination of concept extension and intension.To address problems on the instability of concept extension-based similarity measures and the poor discriminability of concept intension-based similarity measures,from the perspective of the definition of cloud model,based on the relationship between randomness and fuzziness in the generating process of cloud drops,this thesis proposes an uncertainty distribution-based similarity measure of cloud concept,which can reflect the uncertainty distribution of concept extension from the aspect of concept intension.It overcomes the defects of instability and poor discriminability in similarity measures of cloud concept,and realizes the similarity measure based on the combination of concept extension and intension.The experiment results demonstrate that the proposed similarity measure of cloud concept can improve the classification accuracy in the time series classification task,and raises the decision efficiency in the multi-attribute group decision-making process based on the cloud model linguistic variable.(2)Based on fuzziness and imprecision compounded uncertainty form of interval-valued intuitionistic fuzzy(Iv IF)concept,the Iv IF granular structure distances are proposed,which can realize the difference measure of a same concept at various granular layers.In the Iv IF multi-granular space,to address the problem of the inability to accurately distinguish the difference of a same concept at various granular layers based on the current uncertainty measures,this thesis respectively proposes a pair of granular structure distances based on membership and non-membership degrees of Iv IF number to take full account of the hesitant fuzziness and imprecision of numerical value under Iv IF granulating strategy.On this basis,a more widely adapted partial order relation of granular structure,and the granularity measure and information measure defined by this partial order relation are proposed,which realize the difference measure of concept in the multi-granular space.The experiment results demonstrate that this difference measure can not only obtain more simplified reduction sets in decision systems,but also improve the decision efficiency of the multi-attribute group decision-making process based on the Iv IF linguistic variable.(3)In decision system with hierarchical classification(Hie DS),based on the quotient space representation of hierarchical classification(HC),the double granular space structure induced by condition attribute sets and HC decision attribute is proposed.Then,three forms of information entropy are proposed to describe the change rules of granularity layer in double granular space,which can realize the granularity selection mechanism of double granular space.To address the inability of the current approximate measure to reflect hierarchical semantics of target concept,this thesis uses the knowledge distance of object subsets to reflect the difference between the hierarchical semantic labels,and proposes a variety of information entropy to measure the change of inconsistency caused by the transformation of the granularity layers in the double granular space structure.Specifically,the information loss entropy and the described entropy are employed to measure the change of inconsistency caused by the transformation of granularity layers of the HC decision attribute,and the granularity selection mechanism of the HC decision attribute is implemented.The hierarchical information conditional entropy is employed to measure the approximate quality of the condition attribute subset when describing the HC target concept,and a corresponding attribute reduction algorithm is proposed.The experiment results demonstrate that the proposed algorithm can obtain more simplified reduction sets and significantly improve classification accuracy on reduction sets compared with other representative attribute algorithms.This superiority becomes more obvious when the granularity layer is deepening.(4)Based on neighborhood granulating method in the continuous-valued Hie DS,information aggregation functions are constructed,and misclassification and test cost in the approximate describing process for HC target concept are proposed,which can realize the cost-sensitive attribute reduction in continuous-valued Hie DS.To address the inability of information entropies to describe the changes of granularity layers in continuous-valued Hie DS,this thesis respectively proposes the misclassification cost reflecting the distance of labels in HC and test cost reflecting the generalization ability of condition attribute subset.The monotonicity changes of misclassification cost are revealed,and the cost-sensitive attribute reduction in continuous-valued Hie DS is realized.The experiment results demonstrate that the proposed algorithm can obtain more simplified reduction sets and significantly improve classification accuracy on reduction sets. |
