Multi-view Data Classification Research

With the continuous development of scientific and technological information,multi-view data exists widely in real life,which provides richer and more comprehensive information for data analysis.In practice,when a certain amount of labeled samples are obtained,we can use labeled samples to train a classifier,so as to classify unlabeled samples.Due to the high cost of label acquisition,when we only obtain a small number of labeled samples,these can be used together with a large number of unlabeled samples for training,so as to classify unlabeled samples.This process is semi-supervised classification.Therefore,the semi-supervised and supervised classification of multi-view data have been widely used in daily life and gradually become the research hotspots.However,the traditional graph-based semi-supervised classification algorithms only construct the graph from each view,without considering the global structure of the multi-view data,making the classification perfor-mance of the algorithm sub-optimal.At the same time,in multi-view classification,how to obtain the view-consistent representations is the key to the research of multi-view classifi-cation.Most existing methods only use Euclidean distance to measure.The scale-sensitive characteristic of Euclidean distance makes the view-consistent representations obtained by the algorithm are not robust,and the existing methods do not consider the class-specific dis-tribution of different classes of samples.Target at these problems,the research contents of this thesis are as follows:(1)Target at the problem that the graph-based semi-supervised classification algorithms ig-nore the global structure of multi-view data and do not make full use of its rich information,a semi-supervised classification algorithm that combines global and local graphs(SSC-GL)is proposed.We exploit both the global structure embedded in the concatenated views and the local manifold structure embedded in different views.Thus,the learned similarity ma-trix can well characterize the intrinsic relationship between data and obtain a better consis-tency structure.Comprehensive experiments show the effectiveness of proposed approach,especially when the label ratio is small,the algorithm proposed in this thesis has greater performance.(2)Target at the problem that the existing multi-view classification algorithms do not obtain the robust view-consistent representations and do not consider the class-specific distribution of samples,a cross-view classification algorithm by adversarial learning and class-specific distribution(CvALCS)is proposed.Using the idea of GAN combined with Fisher criteri-on to obtain robust view-consistent representations,that is,the view metric learning(Fisher criterion)is used as the generator and the view classifier is used as the discriminator.During the training process,the view metric learning continuously maximizes the probability that the view classifier makes mistakes,and the purpose of the view classifier is to assign labels to the samples as accurately as possible.The adversarial process between the two can better eliminate the gaps between different views and thus obtain robust view-consistent represen-tations.On this basis,a class-specific distribution item measured by the(?)₁₂-norm,makes the view-consistent representations with the same label have a common distribution,while the view-consistent representations with different labels have different distributions in the inherent feature space.We integrate adversarial learning and class-specific distribution into a unified framework,so that the learned view-consistent representations can not only encode and discriminate information well,but also describe the class structure of the dimensional space.Experiments on real data sets prove the effectiveness of the proposed algorithm.