With the advent of the era of big data,machine learning algorithms have been widely used and rapidly developed in multi-view data.The diversity and complementarity of multi-view data not only help machine learning algorithms better discover the essential attributes and pattern partition of data structure,but also bring many challenges to traditional machine learning algorithms.How to analyze and process the high-dimensional,heterogeneous and large-scale multi-view data,so as to deeply explore the hidden information,potential mode and effective knowledge of multi-view data,making it widely used in various fields such as computer vision,pattern recognition,statistical analysis and information retrieval,becomes increasingly important.In view of the complexity of multi-view data,the following aspects need to be paid attention to in its modeling and learning process,such as redundancy reduction of multi-view data,eigen-representation of multi-view data,fusion of multi-view data,and feature processing of multi-view data.Beginning with the construction of appropriate multi-view learning model and algorithm,this paper conducts an in-depth study on multi-view learning algorithm from three aspects including multi-view data clustering,classification and feature selection.The main contributions of this dissertation are summarized as follows:(1)In order to solve the problems of dimensionality reduction,clustering,fusion and eigenrepresentation of multi-view data and to establish a unified multi-view clustering framework,this paper proposes a Discriminatively Embedded K-Means for Multi-view Clustering(DEKM).Firstly,DEKM embeds the dimension reduction technique(i.e.,Linear Discriminant Analysis)into multi-view K-Means clustering framework.With pseudo label information generated by initialization and using multiple linear discriminant analyses,DEKM realizes the synchronous learning of multiple discriminative low-dimensional subspaces via solving the Trace Ratio optimization.Secondly,in the unified multi-view learning framework,DEKM solves the common clustering indicator matrix across different views by minimizing the weighted sum of multiple K-Means clustering objective function values and explores the underlying consensus pattern partition among different views.Finally,DEKM proposes an adaptive weighted method to control the weight assignment by adjusting the power exponent parameter and implements the fusion and intercoordinations among different discriminative low-dimensional subspaces.As an unsupervised multi-view clustering algorithm with dimension reduction,through an alternate optimization method,DEKM not only preserves the view-specific attributes but also maintains the common consistency of pattern partition among different views.(2)In order to enhance the robustness of multi-view clustering algorithm,this paper proposes a Re-weighted Trace Optimization for Multi-view K-Means Clustering(RTMv KM).Firstly,RTMv KM develops the robust data-driven model based on the least-absolute criteria,and transforms original problem into a trace optimization by using the re-weighted least squares.The trace optimization can be solved by eigen decomposition and obtains multiple discriminative low-dimensional subspaces.Secondly,RTMv KM solves the shared clustering indicator matrix across different views by minimizing the weighted sum of multiple K-Means clustering objective function values and explores the underlying consensus pattern partition among different views.Thirdly,RTMv KM adaptively learns multiple weights in a re-weighted manner and each weight only depends on current view-specific projection and cluster indicator,which not only avoids the hyperparameters pre-setting and tuning but also reduces the model complexity.As a robust unsupervised multi-view clustering algorithm integrated dimensionality reduction,RTMv KM can effectively reduce the impact of outliers on multi-view clustering and simplify the preset parameters of the model.(3)Inspired by the ideal similarity matrix containing appropriate neighborhood assignment,in order to construct a sparsity flexible common membership matrix to guide the multi-view clustering,this paper proposes a Multi-view K-Means Clustering with Adaptive Sparse Memberships(Mv ASM).Firstly,Mv ASM introduces a common membership matrix with adaptive sparseness and relaxes the binary clustering indicator vector of each sample to a membership vector whose probabilities are summed up to one,and then adopts a balanced parameter to control the sparseness degree of the membership vectors.Secondly,Mv ASM embeds the adaptive sparse membership matrix into the multi-view K-Means clustering model and utilizes the Newton method to solve the common membership matrix,which explores the underlying consensus pattern partition among various views to guide the multi-view clustering.Finally,through an alternate optimization,Mv ASM learns the clustering centroid matrix and the weight of each view to explore the view-specific clustering structure and integrates the multiple view information to coordinate the complementary information among various views through adaptive weighted assignment.As an unsupervised multi-view fuzzy clustering model,MVASM resorts to the soft clustering partition to effectively alleviate the drawbacks of hard clustering partition.(4)In order to comprehensively consider the deep interactive information and fusion strategies among different views,this paper proposes a Multi-view Neural Networks Embedded Crossview Deep Interactive Information for Categorization(Mv NNcor)to improve the classification performance of multi-view data.Firstly,Mv NNcor captures various intra-view information from different views by multiple sub-networks to learn the intrinsic properties.Outer product is exploited to model the pairwise correlations between the dimensions of the high-level features from each view and all other views,which generates a two-dimensional interaction map that is more expressive.Secondly,based on the generated interaction map of each pairwise views,Mv NNcor further introduces an interaction function to project the vectorized interaction map into an embedding space,which learns cross-view deep interaction information and makes them incorporate with the intra-view information in a proper proportion.Finally,Mv NNcor integrates the above two kinds of information of each view through a deep neural network and calculate its corresponding cross-entropy loss,and then fuse different losses via in an adaptive-weighted way.As a supervised multi-view learning framework,Mv NNcor seamlessly embeds the intraview and inter-view information and builds a new multi-view loss fusion strategy to realize the multi-view collaborative learning and joint decisions during the optimization for categorization,and further improves the classification performance of the multi-view framework.(5)In order to avoid curse of dimensionality,alleviate information redundancy and remove irrelevant features,this paper proposes a Multi-class Scaling Support Vector Machines for Multiview Feature Selection and Classification(denoted as Mv SSVM).Firstly,Mv SSVM embeds the scaling factor of each view into the corresponding multi-class SVM,which can renewedly adjust the weight allocation on all features to make discriminative and important features get larger weights.Secondly,introducing the scaling factors is beneficial to simplify the proposed model which is bridged with the solvable problem through a series of the mathematical deduction,and then can be solved by an alternating iteration optimization method.Finally,Mv SSVM utilizes the view-specific multi-class SVM to obtain the decision values and its corresponding confidence scores and present a novel fusion strategy which calculates the final confidence score of each class by merging the confidence scores of different views,and then determines the category in terms of the highest confidence score among final confidence scores.As a supervised multiview learning framework,MVSL21 uses the common label information cross different views to guide the weight matrix learning and to explore discriminative features for each view,and then jointly infers the category of samples via a fusion strategy of multi-view decision functions. |