Research On Multi-View Construction Model

With the continuous advancement of technology and the continuous improvement of computer performance,artificial intelligence has been developed and penetrated into every aspect of life.Every day,people's clothing,food,housing,transportation,and the normal operation of all walks of life will generate massive amounts of data.How to make full use of these data and explore potential value has become the primary task of machine learning.With the development of data acquisition technology,the data we could obtain is more and more rich,and it is more and more complicated.There are some data expressed in different ways,called multi-view data.Multi-view learning technology is born.Multi-view data has two basic properties,complementarity and consistency,which provide richer information for subsequent machine learning tasks.However,during the data collection process,we may not be able to obtain multiple views of the data,or obtain incomplete multi-view data.How to build multiple views based on existing single view data or construct views using existing multi-view data becomes a valuable topic to research.In practical applications,there are some data with the characteristics of small sample size and high dimension.The traditional machine learning methods are difficult to learn the knowledge well.The dimension reduction algorithm can extract important features,but at the same time lose some information.This thesis proposes a multi-view construction model based on feature set partitioning(FP_MvCM),which combines the traditional dimensionality reduction method with multi-view learning technology to eliminate redundant information while preserving abundant information.In the process of constructing multi-view data,a multi-view quality evaluation method is proposed to obtain high quality multiple views.The constructed multi-view is used for the clustering task,and the validity of the multi-view data is detected by the clustering result.The experimental results show that FP_MvCM guarantees the complementarity and consistency information of multi-view data while removing redundant features.In the process of obtaining data,due to some uncontrollable factors,the multi-view data we get may be incomplete,so how to use these complete multi-view data to fill the incomplete view is a question worth studying.This thesis proposes a multi-view construction model based on self-encoding network(MvAEM),which has an encoder and a decoder.The existing multi-view is input into the encoder to obtain unified coding of multiple views,then decodes the decoder to obtain the target view.The experimental results show that MvAEM can generate a clear target view contour and achieve the twist of the view angle.