With the development of data sensing technology,there are more and more ways to describe the same object,and the features obtained in each way are often called a view.At present,multi-view data exists in the fields of biological science,social network,image recognition,automobile industry and robot industry.In order to make more efficient use of the rich and complementary information contained in multi-view data,the field of machine learning has developed a corresponding multi view learning method.The existing work has shown that if multi-view data is merged into a single view according to specific strategies,the further classification or regression tasks would have better learning effect.This paper studies the feature extraction of multi view data in order to achieve dimensionality reduction and information fusion.From the perspective of Canonical Correlation Analysis(CCA),based on the analysis of the merits and demerits of CCA and its mutation algorithm,the related research is carried out.At present,most multi-view dimensionality reduction algorithms can be categorized as subspace learning in the broad sense.The basic principle of subspace learning is the model of latent variable generation model.It corresponds to the projection model,so CCA belongs to the latter.In addition to the improvement of CCA,a new method of feature extraction and information fusion based on the projection model is proposed.The contributions are as follows:1.Canonical Correlation Analysis is widely used for multiview feature extraction and information fusion since it was proposed in 1936.When used for multiview classification learning,CCA usually extracts features in unsupervised way for the subsequent classification task.As CCA do not use the given discriminant information,the features extracted in this case may be not good for discrimination.For this reason,many kinds of CCA and its variation which can use discriminant information were proposed.The corresponding Objects derived by these methods are usually nonconvex which affects the problem solving and the subsequent classification effect to a certain extent.To our knowledge,there have been no convex discriminate CCA methods were proposed so far.Inspired by Geometric Mean Metric Learning(GMML),we propos a convex discriminate CCA called Convex Discriminant Correlation Analysis(CDCA).CDCA transforms the learning of two projection matrices into a geodesic convex problem of metric learning,thereby admitting a closed form solution and simultaneously extracting discriminant fused features directly.2.There are two kinds of principles in multi-view learning: consistency and complementarity.Consistency comes from the shared information of different views and complementarity comes from the individual information of each view.Most approaches of multi-view dimensionality reduction only focus on either consistency or complementarity.The low dimensional representations learned by those approaches may suffer from information redundancy or information incompleteness.This paper proposes a novel framework of multi-view feature extraction called multiple structured sparsity projection(MSSP),which can extract both shared and individual information adaptively.MSSP is made of two parts: one is a criterion item on features combined with all views' projections,which is implemented by selecting linear discriminant analysis(LDA)and metric learning.The other one is a multiple structured sparsity regularization item on the united projection matrix to ensure the consistency and complementarity.The united projection matrix is composed of all views' projection matrices,and the projection matrix of each view is sparse.So a multiple structured sparse representation is established,so that the shared information and unique information can be extracted adaptively. |