Research And Application Of Incomplete Multi-view Data Representation Learning

In today's Internet world,multi-view data is showing an explosive growth trend.How to efficiently mine the information in multi-view data and serve the public has become one of the many problems in academic research,and the representation learning of multi-view data is also a hot spot.Multi-view data refers to a collection of data that objectively reflects the same object through different perspectives and different forms.Multi-view data has two notable characteristics: consistency and complementarity.Different views often reflect the complementary characteristics of the same object.Therefore,by observing the multi-view data,the characteristics of the object can be more fully understood and more valuable information can be delivered to the user.In addition,multi-view data can also meet more user requirements,such as voice-to-text conversion and image search for video,etc.,providing more possibilities for future technological advancement.These practical applications are closely related to the representation learning of multi-view data.Through representation learning,the data is mapped to a new feature space.The final task only needs to perform simple operations such as calculation and sorting on the new feature space.However,in actual application scenarios,it is difficult to collect complete multi-view data for algorithm learning and use,which may be caused by instrument damage,limitation of collection methods,and other reasons.Incomplete multi-view data includes two forms: 1)multi-view data with missing partial views,2)multi-view data with missing all views(small sample size).However,in actual application scenarios,it is difficult to collect complete multi-view data for algorithm learning and use,which may be caused by instrument damage,limitation of collection methods,and other reasons.Incomplete multi-view data includes two forms: 1)multi-view data with missing partial views,and 2)multi-view data with missing all views(the number of samples is small).At present,most of the research assumes to avoid these two data forms.This assumption limits the application of these algorithms in actual scenarios.In response to the problems raised above,this article focuses on incomplete multi-view data in two different forms.It means learning to study and propose two algorithm models based on two different tasks.The main work results and contributions are as follows:(1)An unsupervised cross-view hashing model(UDIH)based on imputation is proposed,which combines the multi-view data representation learning with missing partial views and cross-view retrieval tasks.The main challenge of cross-view retrieval is that there is a heterogeneous gap between different views.Partial multi-view data will result in a large amount of missing pairwise information(one-to-one relationship between views)and make it more difficult to break the heterogeneous gap.To address these problems,this paper proposes UDIH,which consists of two parts: the first part is to solve the problem of pairwise information loss,we use the dual-channel CRA architecture to generate pairwise information,and the second part uses weighted triple loss and The weighted cross-view relationship graph sets different view data to the public Hamming space through the neural network,and the weighted cross-view relationship graph is constructed based on the expanded information.On the benchmark data set,a large number of experiments verify that UDIH has a better cross-view retrieval performance than the state-of-the-art methods.(2)A prototype-based multi-view discrimination common subspace learning model(PMDSL)is proposed,which combines all missing multi-view data representation learning with multi-view classification task.The main challenge of the multi-view classification task is how to efficiently use consistency and complementary information.Therefore,PMDSL first captures the consistency information between views by restricting the representation of each view in the latent space as close as possible to its corresponding class prototype.Then it retains the complementary information by letting the latent representation inherit the diversity of the multi-view data.Finally,PMDSL adaptively merges multiple discriminant representations into a comprehensive representation by considering the correlation between features.Both latent representation and fusion representation are based on the loss of prototypes by emphasizing the compactness within the class and the separability between classes,which improves the robustness and resolution of the representation.A large number of experimental results on four multi-view benchmark data sets show that PMDSL is superior to existing methods in terms of accuracy and robustness.