Research On Multiview Correlative Representation Learning Algorithms Based On Covariation Projection

With different data collection methods and perspectives of data description,we usually obtain multiple representations of a same object.In the literature,these data are referred to multiview data.In order to fully exploit the associated information among multiview data,some classic multiview representation learning methods employ canonical correlation analysis to extract latent representations.Although these methods can learn useful multiview representations,there still exist some challenging problems in practice for them.For example,they can hardly reveal the complex relationship among feature vectors.Moreover,the singularity problem in small sample cases leads to limited practicality.To address the aforementioned problems,we propose a series of covariation projection based multiview correlative representation learning methods for multiview data.In addition,we verify their effectiveness in pattern classification,clustering,face super-resolution reconstruction and recognition tasks.Main contributions and innovations of this paper can be summarized as follows:(1)We propose a correlative covariation projection(CCP)algorithm.To tackle the problem that conventional canonical correlation methods can hardly discover nonlinear relationship among feature vectors,we construct covariation matrices of feature vectors with Gaussian kernel mapping and propose a novel canonical correlation analysis algorithm based on it,namely CCP.It can not only capture nonlinear relationship among features,but also avoid the singularity problem of matrices in small sample scenarios.Extensive experimental results demonstrate the effectiveness of CCP.(2)We propose a deep covariation projection(DCP)algorithm.To address the issue that CCP method cannot be effectively used in large-scale data learning,we propose DCP algorithm,which utilizes deep neural networks to learn two-view correlative representation.It maximizes the correlation between deep representations based on deep covariation matrices and preserves the structure information of original data through alignment module.Experimental results illustrate that DCP outperforms related methods in various tasks.(3)We propose a multiset covariation projection framework.For the multiset scenario with more than two views,we propose multiset extensions for both CCP and DCP,which are multiset CCP(MCCP)and multiset DCP(MDCP),respectively.MCCP maximizes the sum of correlation between different views to learn multiview correlative representation.It can be solved by an iterative method,whose convergence is proved theoretically.MDCP maximizes the accumulated correlation and minimizes multiset reconstruction error.It can be optimized by stochastic gradient descent algorithm.Results of experiments indicate that MCCP and MDCP are superior to existing representation learning methods.