Subspace-based Multi-view Learning Theory And Method Research

With the rapid growth of information collection techniques,multi-view data is u-biquitous in most real applications.How to extract useful information from multi-view data has become an active research topic in the field of pattern recognition.Different views of the object describes different aspects and complement each other.Therefore,it is significant to determine the latent low-dimensional representation by exploiting the complementary information among views,which contributes to the most profound under-standing of the object,improves subsequent data analysis and reduces the computational complexity simultaneously.From the point of shard subspace learning,this dissertation engages in academic research on complete multi-view data representation learning and incomplete multi-view data representation learning by considering the characteristics of multi-view data,i.e.,high dimensionality and high noise,the correlations and differences among multiple views.The main contributions are summarized as follows:(1)Most of the multi-view data representation learning methods assume that each view data is independent of each other and has the same contribution to the final semantic representation,which ignores the high-order correlation and differences between views,and therefore results in the complementarity between different views can not be effectively exploited.To address this issue,a tensor-based low dimension-al representation learning method(tRLMvC)for multi-view clustering is proposed in this thesis.The model includes two parts:one is for learning multi-view data self-expressive tensor,and the other one is for factorizing self-expressive tensor vi-a Tucker decomposition.To capture the high-order correlation among views and objects,tRLMvC firstly represents multi-view data as a third-order tensor and then employs the t-product based on the circular convolution operation to learn the self-expressive tensor.To determine the low-dimensional representation by considering the differences among views,tRLMvC applies Tucker decomposition on the self-expressive tensor.These two parts are iteratively performed so that the interaction between self-expressive tensor and its factorization can be enhanced.tRLMvC can determine each view's importance and learn the low-dimensional unified repre-sentation of objects by integrating multi-view information suitably.Experiments validate that multi-view representation learning benefits from its merits to includ-ing sufficient mining the high-order correlations among views and considering the differences among views(2)Aiming the problem that the data objects from one view or more than one view may be inaccessible,a novel incomplete cross-view hashing representation learning method(RUCMH)is proposed in this thesis.The major contribution lies in jointly learning modal-specific hash function,exploring the correlations among views with partial even without any pairwise correspondence and preserving the information among original features as much as possible.More specifically,to determine the correlations across different views,RUCMH maps multiple heterogeneous feature spaces to a shared space,in which each object of one view will be reconstructed by the objects in the other view.In this way,the construction coefficients can represent the correlations among objects from different views.Meanwhile,the orthogonal constraint on the maps can keep the information of original features as much as possible.Extensive experiments on fully-paired,partially-paired and even no-paired data settings demonstrate that RUCMH can achieve relative stable cross-modal retrieval performance.(3)To handle the high dimensional and noisy multi-view data,a novel incomplete multi-view data representation learning method(WMvRL)with feature weighting is proposed in this thesis.WMvRL is suitable for all kinds of multi-view data,including partial and complete cases.Its goal is to sufficiently capture the corre-lation among multiple views and objects,quantify the importance of features and the contributions of views so that a more accurate representation can be obtained.More specifically,it can identify the shared low-dimensional feature space from o-riginal multiple heterogeneous feature spaces by embedding feature weighting and re-represent data objects in the latent space.To emphasize the informative features in each view while de-emphasizing uninformative features,WMvRL introduces a variable to weight each feature in the view.To seamlessly capture high-order correlations among multi-view data,WMvRL merges the specific self-expressive matrices of multiple views into a 3-order tensor and then employs the tensor nu-clear norm constraint on the self-expressive tensor.To determine the shared low-dimensional latent space and quantify the contribution of each view simultaneously,WMvRL factorizes the self-expressive tensor via Tucker decomposition.Experi-ments validate that multi-view representation learning benefits from its merits to discriminate the differences among views and the differences among features in each view.