Research On Attribute Reduction And Classification Of Imcomplete Multimodal Data

In big data era,data contains valuable information.By mining the potential characteristics and rules in the data,we can provide reference information for many decisions in real life,which has important research value.These data are massive,incomplete,multimodal and so on.For massive data,an efficient attribute reduction algorithm based on local conditional distinguishing ability is proposed in this thesis.For incomplete multimodal data,an integrated classification method of incomplete multimodal data based on view selection and modal fusion is proposed in this thesis.The main innovations of this paper are as follows:(1)This thesis constructed the conditional distinguishing ability based on the distinguishing relationship,which theoretically proves that the conditional distinguishing ability is more strict and reasonable than the positive field in the selection of attributes,and proposes an attribute reduction algorithm based on the conditional distinguishing ability.In order to speed up the calculation of attribute importance and improve the efficiency of reduction,this thesis extends the conditional distinguishing ability to the local conditional distinguishing ability by random sampling according to the stability of frequency in the law of large numbers.An attribute reduction algorithm based on local conditional distinguishing ability is further proposed,and the temporal and spatial complexity of the algorithm is analyzed.Finally,comparative experiments show that the proposed algorithm has higher reduction efficiency and is suitable for attribute reduction of massive data.(2)Aiming at the characteristics of incomplete multimodal data,this thesis uses mutual information to measure attribute correlation,constructs attribute correlation graph,designs a recursive solution method for all maximal cliques of graph,and then proposes a new view selection method.Using view combination to map multiple complete single modal data subsets from the source data,effectively reducing the information redundancy within and between views,and greatly covering source data information.The attribute reduction of each view is carried out to further reduce the information redundancy in the view.Then the basic classifier is trained for each view,and the final decision is made by weighted voting method.Finally,combine with the two modal fusion methods of data level fusion and decision level fusion,an incomplete multimodal data integration classification model based on view selection and modal fusion is proposed.The experimental results show that the integrated classification model proposed in this thesis has good classification effect.